Interest modeling method and system for browsing behaviors based on big data

Abstract

The invention provides an interest modeling method and system for browsing behaviors based on big data, and the method comprises the steps: carrying out the feature fusion according to collected listening habit data, user reading annotation data and voiceprint data, and forming a mixed feature space, dividing the users into different clusters by using K-means, each cluster representing a user type, and obtaining a comprehensive feature vector representing user interests and features; establishing a co-occurrence graph, performing topic modeling by using LDA, integrating popularity, performing normalization, updating a model, calculating a popularity time sequence and predicting future popularity; the change trend of content popularity is researched according to the time sequence characteristics, the use characteristics of listening content data, reading data and voiceprint data of the user are analyzed, and the user preference is mined by adopting a multi-modal method. The accuracy and diversity of recommendation are improved, and the satisfaction degree and the use experience of the user are guaranteed.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to an interest modeling method and system based on big data browsing behavior. Background Technology

[0002] With the development of information technology in modern society, people are increasingly reliant on computer technology for accessing and using information. On the internet, people can easily access various types of information, including news, music, videos, pictures, and more. However, due to the sheer volume of information, people often need to spend a significant amount of time and energy searching for information that interests them, which causes considerable inconvenience in information access and use. Therefore, how to provide personalized information services based on users' needs and interests is a crucial issue in the current development of the internet.

[0003] To address this issue, a method and system for interest modeling based on big data browsing behavior is proposed to provide personalized information services. This solution uses big data browsing behavior for interest modeling, utilizes hashing technology to hash voiceprint data, and uses "stickiness" and "response rate" formulas to determine content attractiveness, increase diversity and recommendations for emerging content, update user interest models, and improve the accuracy and diversity of recommendations. Summary of the Invention

[0004] The purpose of this invention is to address the problem that, due to the vast amount of information available, people often need to spend a lot of time and energy to find information that interests them, which causes great inconvenience to people's information acquisition and use.

[0005] To achieve the above-mentioned objectives, this invention provides an interest modeling method and system based on big data browsing behavior to improve the aforementioned problems.

[0006] The application is as follows: An interest modeling method based on big data browsing behavior includes: S110. Based on the collected listening habit data, user reading annotation data and voiceprint data, feature fusion is performed to form a hybrid feature space. K-means is used to divide users into different clusters, each cluster representing a user type, to obtain a comprehensive feature vector representing user interests and characteristics. S120, Build a co-occurrence graph, use LDA for topic modeling, integrate popularity, normalize, update the model, calculate the popularity time series and predict future popularity; S130. Study the changing trend of content popularity based on time series characteristics, analyze the usage characteristics of users' listening content data, reading data and voiceprint data, and use multimodal methods to mine user preferences; S140. Construct a recommendation content model using generative models and seq2seq models; S150. Use user listening habit data, user reading annotation data and voiceprint data to train the GRU4Rec sequence recommendation model. Use transfer learning to pre-train the GRU4Rec model to make up for the lack of data and build the GRU4Rec sequence recommendation model. S160. Hash the voiceprint data using hashing technology SHA, collect listener habits and reading annotation data, use the "stickiness" and "response rate" formulas to determine the attractiveness of the content, optimize the recommended content model based on the feedback results, increase the recommendation of emerging and diverse content, update the user interest model, and use the seq2seq model to update the GRU4Rec sequence recommendation model and recommended content model to improve accuracy. S170. Sparse Bayesian learning is used to establish the mapping relationship between voiceprint features and emotional states, and the emotional state corresponding to the input voiceprint data is determined by calculating the posterior probability of each emotional state.

[0007] An interest modeling system based on big data browsing behavior includes: The popularity analysis module is used to select content with higher popularity for recommendation; The path analysis and association rule analysis modules for listening content data, reading data, and voiceprint data are used to analyze users' listening habits, reading annotation data, and voiceprint data, as well as the association rules between them, to provide a foundation for the content recommendation model. The attention mechanism module is used to enhance the focus on user interests and improve the accuracy of recommendations; The content category clustering module is used to classify content, providing a foundation for the content recommendation model; The Recommendation Content Model module is used to build a recommendation content model and generate personalized recommendation content. The GRU4Rec sequence recommendation model module is used to train the GRU4Rec sequence recommendation model based on user listening habit data, user reading annotation data, and voiceprint data, and to make personalized recommendations. A voiceprint data hashing module based on hashing technology SHA is used to hash voiceprint data to ensure the security of user information. The "stickiness" and "response rate" formulas are used to determine content attractiveness, optimize the content recommendation model, increase diversity, and recommend emerging content. The module for using diversity metrics is used to increase the diversity of recommended content; The module used in the SVM model is used to update the user interest model; The module for establishing a mapping relationship between sparse Bayesian learning and emotional state is used to establish the mapping relationship between voiceprint features and emotional state and to determine the emotional state corresponding to the input voiceprint data.

[0008] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: In the scheme of this application: 1. More accurate interest modeling, using multiple data sources for training, including user audience habit data, user reading annotation data and voiceprint data, to gain a more comprehensive understanding of user interests and needs; 2. More personalized recommendations: The GRU4Rec sequence recommendation model is trained on user audience habit data, user reading annotation data, and voiceprint data, which can more accurately recommend content that users are interested in. 3. More secure voiceprint data processing: Hash technology is used to hash voiceprint data to ensure the security and privacy protection of user information; 4. By employing sparse Bayesian learning to establish a mapping relationship between voiceprint features and emotional states, this approach can more accurately determine the user's emotional state, thereby improving user experience and recommendation accuracy, and making the recommendation system more closely aligned with the user's actual needs and interests. Attached Figure Description

[0010] Figure 1 The flowchart illustrating an interest modeling method based on big data browsing behavior according to an embodiment of this application is shown schematically. Figure 2 A block diagram of an interest modeling system based on big data browsing behavior according to an embodiment of this application is illustrated schematically. Detailed Implementation

[0011] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0012] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0013] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0014] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0015] Example 1 The implementation details of the technical solutions in the embodiments of this application are described in detail below: Figure 1 A flowchart of an interest modeling method based on big data browsing behavior according to an embodiment of this application is shown. The method includes at least steps S110 to S170, which are described in detail below: In step S110, feature fusion is performed based on the collected listening habit data, user reading annotation data and voiceprint data to form a hybrid feature space. Then, K-means is used to divide users into different clusters, with each cluster representing a user type, to obtain a comprehensive feature vector representing user interests and characteristics. In one embodiment of this application, feature fusion is performed based on collected listening habit data, user reading annotation data, and voiceprint data to form a hybrid feature space. K-means is then used to divide users into different clusters, each cluster representing a user type. Finally, a comprehensive feature vector representing user interests and characteristics is obtained, including: The filtering method Y=WX, the wrapping method Y=[WTX;X], and the embedding method Z=φ(X) combine the collected listening habit data, user reading annotation data, and voiceprint data to form a hybrid feature space. Continuous feature normalization and discrete feature one-hot encoding are used to transform them into the same feature space. The K-means clustering algorithm is used to divide users into different clusters, with each cluster representing a user type. The interest model of the corresponding cluster is constructed based on the distribution of the user's data feature space within each cluster. For each user type, a comprehensive feature vector representing interests and features is obtained.

[0016] In step S120, a co-occurrence graph is created, topic modeling is performed using LDA, popularity is integrated, normalization is performed, the model is updated, the popularity time series is calculated, and future popularity is predicted. In one embodiment of this application, and in some embodiments thereof, based on the foregoing scheme, a co-occurrence graph is established, topic modeling is performed using LDA, popularity is integrated, normalization is performed, the model is updated, popularity time series is calculated, and future popularity is predicted, including: The content popularity is updated based on the co-occurrence graph using a random walk algorithm. The co-occurrence graph between contents is constructed using listening habit data, user reading annotation data, and voiceprint data. Nodes represent contents, and edges represent the association or similarity between contents, so that the popularity can comprehensively consider the association between contents and reflect the popularity of the contents. Using LDA to model content by topic, the popularity of each topic is calculated as follows: Topic popularity = Sum of popularity of all content within the topic / Total number of content within the topic; The formula for calculating overall popularity is: Overall popularity = Sum of content popularity + Sum of topic popularity; The popularity of content, the popularity of the theme, and the overall popularity are normalized to 0-1 to eliminate the influence of dimensions and make the popularity comparable. The popularity calculation model is updated based on new data and comprehensive feature vectors representing user interests, and the popularity of each dimension is recalculated to track the dynamic changes in popularity. Based on the formula for the change of popularity and the comprehensive feature vector, the time series is calculated to study the trend of popularity over time, and the time series is fitted using the ARMA model. The ARMA model is dynamically updated, and its parameters are recalculated periodically based on new data and comprehensive feature vectors. Existing popularity time series data is used as input, and then the popularity in the future time period is predicted based on the parameters of the ARMA model and historical data.

[0017] In one embodiment of this application, the ARMA model comprises two parts: an autoregressive (AR) model and a moving average (MA) model. The autoregressive part describes the relationship between the current value and several past values, and the moving average part describes the relationship between the current value and several past error values. ARMA model: y_t=c+Σ(φ_i*y_(ti))+Σ(θ_j*ε_(tj))+ε_t Where y_t is the value at time t, p and q are the orders of AR and MA, φ_i and θ_j are the parameters of the model, c is a constant, and ε_t is the error term; the error term ε_t is usually a normally distributed random variable with a mean of 0 and a variance of σ^2. An ARMA model is trained using existing popularity time series data to obtain the model parameters φ_i and θ_j. Then, the parameters are recalculated based on new data and comprehensive feature vectors to obtain new parameter values. Finally, the popularity in future time periods is predicted using the new parameter values ​​and historical data.

[0018] In step S130, the changing trend of content popularity is studied based on time series characteristics, the usage characteristics of users' listening content data, reading data and voiceprint data are analyzed, and a multimodal method is used to mine user preferences; In one embodiment of this application, the study analyzes the changing trends of content popularity based on time-series characteristics, analyzes the usage characteristics of users' listening content data, reading data, and voiceprint data, and employs a multimodal approach to mine user preferences, including: Based on the predicted future prevalence, select the time period with high prevalence and an upward trend for analysis; By using path analysis and association rules to analyze users’ listening content data, reading data and voiceprint data, the three data types are encoded into feature vectors X1, X2 and X3 respectively. The usage frequency and usage path of each data type are counted, and the association rules between data types are determined. An attention mechanism is used to input feature vectors X1, X2 and X3 of three data types in parallel, learn the weight of each feature vector, and obtain the user feature vector AU. This vector comprehensively considers different data sources to determine the user's possible preferences. The K-means algorithm is used to cluster the content categories to obtain a relatively general set of preference categories C that users are interested in; the similarity between AU and each category is calculated, and the top N categories with higher similarity values ​​are determined as the user's preference categories C'. By comprehensively utilizing prediction results, path analysis results, association rule analysis results, and multimodal analysis results, we can determine the popularity of content and user data usage characteristics within each time period, and mine user preference category C'.

[0019] In step S140, a recommendation content model is constructed using a generative model and a seq2seq model; In one embodiment of this application, a recommendation content model is constructed using a generative model and a seq2seq model, including: Using preference category C′ as the training label, and taking the user's listening content feature vector X1 and voiceprint data feature vector X2 as input, a generative model and a seq2seq model are trained. The generative model is used to determine the matching degree between the user's data feature vector and each category, and to infer the content that the user is interested in; y_gen=GenerateModel(X1,X2,C′), the seq2seq model encodes the user's voiceprint data sequence S to obtain the vector encodings d_u and d_c of the user's voice features, d_u,d_c=seq2seq(S); the y_gen, voice encoding d_c and the original sequence S produced by the generative model are input into the recommendation content generation function h() to construct the recommendation content model and generate personalized recommendation content y, y=h(y_gen,S,d_c).

[0020] In step S150, the GRU4Rec sequence recommendation model is trained using user listening habit data, user reading annotation data, and voiceprint data. The problem of insufficient data is made up by using transfer learning to pre-train the GRU4Rec model, and the GRU4Rec sequence recommendation model is constructed. In one embodiment of this application, a GRU4Rec sequence recommendation model is trained using user listening habit data, user reading annotation data, and voiceprint data. Transfer learning is used to pre-train the GRU4Rec model to compensate for insufficient data. The constructed GRU4Rec sequence recommendation model includes: A GRU4Rec sequence recommendation model is trained using user listening habit data, user reading annotation data, and voiceprint data: h = GRU(x,h), where x represents the current input, which is one of the three types of user data; h represents the hidden state; and h' represents the updated hidden state. Data compensation is achieved by pre-training the GRU4Rec model using transfer learning: users with three types of data highly similar to the target user are selected from a large number of users; the GRU4Rec model is pre-trained using the three types of data from the selected users; the weights of the first two hidden layers of the GRU4Rec model are fixed, and the third hidden layer and output layer are fine-tuned using the three types of data from the target user. Personalized recommendations are then made using the GRU4Rec sequence recommendation model based on the user's listening habit data and reading annotation data.

[0021] Specifically, the third hidden layer and output layer are fine-tuned using three types of data from the target user in the GRU4Rec model: The target user's listening habit data, user reading annotation data, and voiceprint data are input into the pre-trained GRU4Rec model to obtain the outputs and weights of the first two hidden layers; Using the outputs of the first two hidden layers as input, continue forward propagation to obtain the output of the third hidden layer; The output of the last hidden layer is used as the input of the softmax layer. Backpropagation is performed based on the target user's behavior sequence to update the weights of the third hidden layer and the softmax layer.

[0022] In step S160, the voiceprint data is hashed using the hashing technique SHA, listener habits and reading annotation data are collected, the "stickiness" and "response rate" formulas are used to determine the attractiveness of the content, the recommendation content model is optimized based on the feedback results, the recommendation of emerging and diverse content is increased, the user interest model is updated, and the GRU4Rec sequence recommendation model and the recommendation content model are updated using the seq2seq model to improve accuracy. In one embodiment of this application, the voiceprint data is hashed using the hashing technique SHA, listener habits and reading annotation data are collected, and the "stickiness" and "response rate" formulas are used to determine the content attractiveness. Based on the feedback results, the recommended content model is optimized, the recommendation of emerging and diverse content is increased, the user interest model is updated, and the GRU4Rec sequence recommendation model and recommended content model are updated using a seq2seq model to improve accuracy. This includes: The newly acquired voiceprint data is hashed using the hashing technique SHA, and the corresponding hash value is queried in the user voiceprint database to determine the user's identity. Collect updated audience habit data and use the "stickiness" formula to determine content attractiveness: Stickiness = Interval Time / Total Time × 100%; collect user reading annotation data and use the "response rate" formula to judge the attractiveness of recommended content y generated by the recommendation content model: Response Rate = Number of Annotations / Number of Views × 100%; Based on the feedback analysis of y, optimize the recommendation content model by increasing or decreasing the number of y items to improve user experience. Use accuracy and recall as metrics to precisely adjust the number of recommended content items y: AOD=∑ni=1ri / n, R=∑ni=1ri / t; recommend content that users have not yet seen but may be interested in to optimize the diversity of recommended content y, using the diversity metric of recommended content y: Diversity=1 / ∑ni=1P2(i); increase the recommendation of emerging content and long-tail content to meet the interests of more users.

[0023] Update the user interest model: Update the user preference category C' using the K-means algorithm, Ck=argminck∑(x-μk)2; Use the SVM model and update the feature vector AU using maxa,b{∑ni=1αi−(1 / 2)∑ni,j=1αiαjyiyj(xi·xj)}; where αi≥0 are Lagrange multipliers used to satisfy the constraint ∑ni=1αiyi=0; x is the recommended content y generated by the recommendation model; y is the label of x, with a value of ±1; ∑ni,j=1αiαjyiyj(xi·xj) is the Errors term, and the separating hyperplane is obtained by maximizing the margin, thereby updating the user interest classification; The seq2seq model used maximizes the log-likelihood, l(θ)=∏n=1p(yn|xn,θ), to update the GRU4Rec sequence recommendation model and the recommendation content model, thereby improving the accuracy of the model.

[0024] In step S170, sparse Bayesian learning is used to establish the mapping relationship between voiceprint features and emotional states, and the emotional state corresponding to the input voiceprint data is determined by calculating the posterior probability of each emotional state.

[0025] In one embodiment of this application, sparse Bayesian learning is used to establish a mapping relationship between voiceprint features and emotional states, and the emotional state corresponding to the input voiceprint data is determined by calculating the posterior probability of each emotional state, including: Sparse Bayesian learning (SBL) is used to establish the mapping relationship between voiceprint features and emotional states, and to determine the emotional state corresponding to the input voiceprint data: p(t|y,X)=∫p(y|z)p(z|X,t)p(z)dz, where t represents the emotional state, y represents the observed voiceprint features, and z represents the hidden voiceprint representation, in order to determine the emotional state corresponding to the input voiceprint data; Sparse Bayesian learning (SBL) is used to establish the mapping relationship between voiceprint features and emotional states, and to determine the emotional state corresponding to the input voiceprint data: p(t|y,X)=∫p(y|z)p(z|X,t)p(z)dz, where t represents the emotional state, "positive", "neutral", and "negative"; y represents the observed voiceprint features, including energy and rhythm features of each frequency band; z represents the hidden voiceprint representation, which needs to be learned; X represents the given training speech data; and p(y|z) represents the conditional probability of y given z, indicating the generation process of y. p(z|X,t) represents the posterior probability of z given training data X and emotional state t; p(z) represents the prior probability of z; ∫ node represents the summation of all z; the optimal representation of z is obtained by maximizing log p(t|y,X), the mapping relationship between z and input voiceprint y is established, and the emotional state t corresponding to the new input voiceprint y is determined accordingly. Given new input voiceprint data y, calculate log p(t|y,X) for each emotional state t, and select the emotional state with the highest log probability as the emotional state corresponding to y.

[0026] In practical applications, a seq2seq model is adopted, and the log-likelihood under the condition of maximization is used to update the GRU4Rec sequence recommendation model and the recommendation content model: Using preference category C' as the training label, and taking the user's listening content feature vector X1 and voiceprint data feature vector X2 as input, a generative model and a seq2seq model are trained. The generative model is used to determine the matching degree between the user's data feature vector and each category, thereby inferring the content that the user is interested in.

[0027] The user's voiceprint data sequence S is encoded, and the seq2seq model is used to obtain the vector encodings d_u and d_c of the user's voice features.

[0028] The generator model outputs y_gen, the audio encoding d_c, and the original sequence S as inputs. The recommended content model is constructed through the recommended content generation function h(), thereby generating personalized recommended content y.

[0029] We employ a seq2seq model and utilize the log-likelihood under maximization conditions to update the GRU4Rec sequence recommendation model and the recommendation content model.

[0030] The above approach establishes a mapping relationship between voiceprint features and emotional states by employing sparse Bayesian learning, encodes the user's voiceprint data sequence using a seq2seq model, and combines a generative model, a seq2seq model, and a recommendation content generation function to generate personalized recommendation content. By maximizing the log-likelihood under certain conditions, the GRU4Rec sequence recommendation model and the recommendation content model are updated. This approach not only improves the accuracy of emotional state judgment but also captures the user's interests and preferences, thereby enhancing the accuracy and diversity of recommendations and ensuring user satisfaction and user experience.

[0031] Example 2 The following describes an apparatus embodiment of this application, which can be used to execute the interest modeling method based on big data browsing behavior in the above embodiments of this application. It is understood that the apparatus can be a computer program (including program code) running on a computer device, such as an application software. The apparatus can be used to execute the corresponding steps in the method provided in the embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the above embodiments of the interest modeling method based on big data browsing behavior in this application.

[0032] Figure 2A block diagram of an interest modeling system based on big data browsing behavior according to an embodiment of this application is shown.

[0033] Reference Figure 2 As shown, an interest modeling system 200 based on big data browsing behavior according to an embodiment of this application includes: The popularity analysis selection module 201 is used to select content with higher popularity for recommendation; The path analysis and association rule analysis module 202 for listening content data, reading data, and voiceprint data is used to analyze users' listening habits, reading annotation data, and voiceprint data, as well as the association rules between them, to provide a foundation for the content recommendation model.

[0034] The attention mechanism module 203 is used to enhance attention to user interests and improve the accuracy of recommendations; The content category clustering module 204 is used to classify content, providing a foundation for the content recommendation model; The recommendation content model module 205 is used to build a recommendation content model and generate personalized recommendation content. The GRU4Rec sequence recommendation model module 206 is used to train the GRU4Rec sequence recommendation model based on user listening habit data, user reading annotation data, and voiceprint data, and to make personalized recommendations. The voiceprint data hashing module 207 based on hashing technology SHA is used to hash voiceprint data to ensure the security of user information. The "stickiness" and "response rate" formulas are used in module 208 to determine content attractiveness, optimize the recommended content model, increase diversity, and recommend emerging content. Module 209, which uses diversity metrics, is used to increase the diversity of recommended content; SVM model usage module 210 is used to update the user interest model; The sparse Bayesian learning module 211 establishes a mapping relationship between voiceprint features and emotional states, and is used to establish a mapping relationship between voiceprint features and emotional states to determine the emotional state corresponding to the input voiceprint data.

[0035] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0036] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor; wherein, the names of these units do not constitute a limitation on the unit itself under certain circumstances.

[0037] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations described above.

[0038] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0039] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0040] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0041] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An interest modeling method based on big data browsing behavior, characterized in that, include: S110. Based on the collected listening habit data, user reading annotation data and voiceprint data, feature fusion is performed to form a hybrid feature space. K-means is used to divide users into different clusters, each cluster representing a user type, to obtain a comprehensive feature vector representing user interests and characteristics. S120, Build a co-occurrence graph, use LDA for topic modeling, integrate popularity, normalize, update the model, calculate the popularity time series and predict future popularity; S130. Study the changing trend of content popularity based on time series characteristics, analyze the usage characteristics of users' listening content data, reading data and voiceprint data, and use multimodal methods to mine user preferences; S140. Construct a recommendation content model using generative models and seq2seq models; S150. Use user listening habit data, user reading annotation data and voiceprint data to train the GRU4Rec sequence recommendation model and build the GRU4Rec sequence recommendation model. S160. Hash the voiceprint data using hashing technology SHA, collect listener habits and reading annotation data, use the "stickiness" and "response rate" formulas to determine the attractiveness of the content, optimize the recommended content model based on the feedback results, increase the recommendation of emerging and diverse content, update the user interest model, and use the seq2seq model to update the GRU4Rec sequence recommendation model and the recommended content model. S170. Sparse Bayesian learning is used to establish the mapping relationship between voiceprint features and emotional states, and the emotional state corresponding to the input voiceprint data is determined by calculating the posterior probability of each emotional state.

2. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S110 includes: The selection method Y=WX, the wrapping method Y=[WTX;X], and the embedding method Z=φ(X) combine the collected listening habit data, user reading annotation data, and voiceprint data to form a hybrid feature space; Continuous feature normalization and discrete feature one-hot encoding are used to transform them into the same feature space; Users are divided into different clusters using the K-means clustering algorithm, with each cluster representing a user type. Construct the interest model of the corresponding cluster based on the distribution of user data feature space within each cluster; For each user type, a comprehensive feature vector representing interests and characteristics is obtained.

3. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S120 includes: The content popularity is updated based on the co-occurrence graph using a random walk algorithm. The co-occurrence graph between contents is constructed using listening habit data, user reading annotation data, and voiceprint data. Nodes represent contents, and edges represent the association or similarity between contents, so that the popularity can comprehensively consider the association between contents and reflect the popularity of the contents. Using LDA to model content by topic, the popularity of each topic is calculated as follows: Topic popularity = Sum of popularity of all content within the topic / Total number of content within the topic; The formula for calculating overall popularity is: Overall popularity = Sum of content popularity + Sum of topic popularity; The popularity of content, the popularity of the theme, and the overall popularity are normalized to 0-1 to eliminate the influence of dimensions and make the popularity comparable. The popularity calculation model is updated based on new data and comprehensive feature vectors representing user interests, and the popularity of each dimension is recalculated to track the dynamic changes in popularity. Based on the formula for the change of popularity and the comprehensive feature vector, the time series is calculated to study the trend of popularity over time, and the time series is fitted using the ARMA model. The ARMA model is dynamically updated, and its parameters are recalculated periodically based on new data and comprehensive feature vectors. Existing popularity time series data is used as input, and then the popularity in the future time period is predicted based on the parameters of the ARMA model and historical data.

4. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S130 includes: Based on the predicted future prevalence, select the time period with high prevalence and an upward trend for analysis; By using path analysis and association rules to analyze users’ listening content data, reading data and voiceprint data, the three data types are encoded into feature vectors X1, X2 and X3 respectively. The usage frequency and usage path of each data type are counted, and the association rules between data types are determined. An attention mechanism is used to input feature vectors X1, X2 and X3 of three data types in parallel, learn the weight of each feature vector, and obtain the user feature vector AU. This vector comprehensively considers different data sources to determine the user's possible preferences. The K-means algorithm is used to cluster the content categories to obtain a relatively general set of preference categories C that users are interested in; the similarity between AU and each category is calculated, and the top N categories with higher similarity values ​​are determined as the user's preference categories C'. By comprehensively utilizing prediction results, path analysis results, association rule analysis results, and multimodal analysis results, we can determine the popularity of content and user data usage characteristics within each time period, and mine user preference category C'.

5. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S140 includes: Using preference category C′ as the training label, and taking the user's listening content feature vector X1 and voiceprint data feature vector X2 as input, the generative model and seq2seq model are trained. Generative models are used to determine the degree of matching between a user's data feature vector and each category, and to infer the content that the user is interested in; y_gen=GenerateModel(X1,X2,C′), the seq2seq model encodes the user's voiceprint data sequence S to obtain the vector encodings d_u and d_c of the user's voice features, d_u,d_c=seq2seq(S); Input the generated model's output y_gen, sound encoding d_c, and original sequence S into the recommendation content generation function h() to construct the recommendation content model and generate personalized recommendation content y, y=h(y_gen,S,d_c).

6. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S150 includes: The GRU4Rec sequence recommendation model is trained using user listening habit data, user reading annotation data, and voiceprint data: h=GRU(x,h), where x represents the current input, which is one of the three types of user data; h represents the hidden state; and h' represents the updated hidden state. The data was supplemented by using transfer learning to pre-train a GRU4Rec model: Select three categories of users whose data is highly similar to that of the target users from a large user base. The GRU4Rec model was pre-trained using three types of data from selected users; By fixing the weights of the first two hidden layers of the GRU4Rec model, the third hidden layer and the output layer are fine-tuned using three types of data from the target user. The GRU4Rec sequence recommendation model is then used to make personalized recommendations based on the user's audience habit data and reading annotation data.

7. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S160 includes: The newly acquired voiceprint data is hashed using the hashing technique SHA, and the corresponding hash value is queried in the user voiceprint database to determine the user's identity. Collect updated audience habit data and use the "stickiness" formula to determine content attractiveness: Stickiness = Interval Time / Total Time × 100%; collect user reading annotation data and use the "response rate" formula to judge the attractiveness of recommended content y generated by the recommendation content model: Response Rate = Number of Annotations / Number of Views × 100%; Based on the feedback analysis results of y, optimize the recommended content model, increase or decrease the number of y, improve user experience, and use accuracy and recall as metrics to precisely adjust the number of recommended content y: AOD=∑ni=1ri / n, R=∑ni=1ri / t; Recommend content that users may not have seen but is potentially interested in, optimize the diversity of recommended content y, and adopt the diversity metric of recommended content y: Diversity=1 / ∑ni=1P2(i); increase the recommendation of emerging content and long-tail content to meet the interests of more users; Update the user interest model: Update the user preference category C' using the K-means algorithm, Ck=argminck∑(x-μk)2; Use the SVM model and update the feature vector AU using maxa,b{∑ni=1αi−(1 / 2)∑ni,j=1αiαjyiyj(xi·xj)}; where αi≥0 are Lagrange multipliers used to satisfy the constraint ∑ni=1αiyi=0; x is the recommended content y generated by the recommendation model; y is the label of x, with a value of ±1; ∑ni,j=1αiαjyiyj(xi·xj) is the Errors term, which is obtained by maximizing the margin to find the separating hyperplane, which is used to update the user interest classification; The seq2seq model is adopted, and the log-likelihood under the maximization condition, l(θ)=∏n=1p(yn|xn,θ), is used to update the GRU4Rec sequence recommendation model and the recommendation content model.

8. The interest modeling method based on big data browsing behavior according to claim 1, characterized in that, S170 includes: Sparse Bayesian learning (SBL) is used to establish the mapping relationship between voiceprint features and emotional states, and to determine the emotional state corresponding to the input voiceprint data: p(t|y,X)=∫p(y|z)p(z|X,t)p(z)dz, where t represents the emotional state, y represents the observed voiceprint features, and z represents the hidden voiceprint representation, in order to determine the emotional state corresponding to the input voiceprint data; Sparse Bayesian learning (SBL) is used to establish the mapping relationship between voiceprint features and emotional states, and to determine the emotional state corresponding to the input voiceprint data: p(t|y,X)=∫p(y|z)p(z|X,t)p(z)dz, where t represents the emotional state, "positive", "neutral", or "negative"; y represents the observed voiceprint features, including energy and rhythm features in each frequency band; z represents the hidden voiceprint representation, which needs to be learned; X represents the given training speech data; and p(y|z) represents the conditional probability of y given z, indicating the generation process of y. p(z|X,t) represents the posterior probability of z given training data X and emotional state t; p(z) represents the prior probability of z; ∫ node represents the summation of all z; the optimal representation of z is obtained by maximizing log p(t|y,X), the mapping relationship between z and input voiceprint y is established, and the emotional state t corresponding to the new input voiceprint y is determined accordingly. Given new input voiceprint data y, calculate log p(t|y,X) for each emotional state t, and select the emotional state with the highest log probability as the emotional state corresponding to y.

9. An interest modeling system based on big data user browsing behavior, characterized in that, include: The popularity analysis module is used to select content with higher popularity for recommendation; The path analysis and association rule analysis modules for listening content data, reading data, and voiceprint data are used to analyze users' listening habits, reading annotation data, and voiceprint data, as well as the association rules between them, to provide a foundation for the content recommendation model. The attention mechanism module is used to enhance the focus on user interests; The content category clustering module is used to classify content, providing a foundation for the content recommendation model; The Recommendation Content Model module is used to build a recommendation content model and generate personalized recommendation content. The GRU4Rec sequence recommendation model module is used to train the GRU4Rec sequence recommendation model based on user listening habit data, user reading annotation data, and voiceprint data, and to make personalized recommendations. A voiceprint data hashing module based on hashing technology SHA is used to hash voiceprint data; The "stickiness" and "response rate" formulas are used to determine content attractiveness, optimize the content recommendation model, increase diversity, and recommend emerging content. The module for using diversity metrics is used to increase the diversity of recommended content; The module used in the SVM model is used to update the user interest model; The module for establishing a mapping relationship between sparse Bayesian learning and emotional state is used to establish the mapping relationship between voiceprint features and emotional state and to determine the emotional state corresponding to the input voiceprint data.

Images