Intelligent recommendation method for video content of iptv set-top box based on deep learning

Abstract

The application discloses an IPTV set top box video content intelligent recommendation method based on deep learning, which comprises data collection, data optimization processing, personalized video content recommendation, recommended video content optimization sorting and video content intelligent recommendation. The application relates to the technical field of digital data processing, in particular to an IPTV set top box video content intelligent recommendation method based on deep learning. The method innovatively combines user preferences and business constraints to realize the balance between personalization and business needs. The method innovatively proposes a multi-view user video interaction graph construction, a double-layer attention mechanism and a multi-level contrast loss function design method, so that more accurate personalized video content recommendation is realized. An optimization target function is designed, a population role differentiation search strategy and a dynamic inertia weight are used to improve the optimization algorithm, the optimization of recommended video content sorting is realized, and the effect of video content intelligent recommendation is significantly improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of digital data processing, in particular to an IPTV set-top box video content intelligent recommendation method based on deep learning. BACKGROUND

[0002] The IPTV set-top box video content intelligent recommendation method based on deep learning refers to a kind of deep learning technology, which is used for intelligent analysis and modeling of multi-source data collected by an IPTV set-top box, and outputs personalized video recommendation results meeting user interests and scene requirements, realizes recommendation services with individuality, diversity and real-time, and improves content matching precision and user viewing stickiness.

[0003] However, in the traditional IPTV set-top box video content recommendation method, only user preferences or business constraints are considered, which leads to the imbalance between user preferences and business requirements in the recommendation results, and causes the technical problems of inaccurate recommendation effect and poor user experience. In the existing personalized video content recommendation model, there is a technical problem that it is difficult to fully mine the complex relationship between users and video content, which leads to the limitation of the recommendation effect and the inability to accurately match user interests. In the existing video content sorting method, there is a technical problem of limitation of optimization algorithm, which leads to poor video content sorting effect, thereby affecting the intelligent recommendation effect of video content. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides an IPTV set-top box video content intelligent recommendation method based on deep learning, which is aimed at the technical problem that in the traditional IPTV set-top box video content recommendation method, only user preferences or business constraints are considered, resulting in that the recommended results cannot balance the user preferences and business needs, thereby causing inaccurate recommendation results and poor user experience. The present application innovatively proposes personalized video content recommendation based on user preferences and recommended video content optimization sorting based on business constraints. By simultaneously considering user preferences and business constraints, the accuracy of the recommended results is effectively improved, the user experience is enhanced, the balance between personalization and business needs is achieved, the intelligent recommendation effect of IPTV video content is significantly improved. In view of the technical problem that in the existing personalized video content recommendation model, it is difficult to comprehensively mine the complex relationship between users and video content, resulting in limited recommendation effect and inability to accurately match user interests, the present application innovatively proposes a combination method of multi-view user video interaction graph construction, double-layer attention mechanism and multi-level contrast loss function design. By multi-view user video interaction graph construction, the potential connection between user behavior and video content can be more comprehensively captured. By double-layer attention mechanism, user multi-dimensional and multi-perspective interests are considered, improving the personalization and accuracy of the recommendation system. By designing multi-level contrast loss function, combining local and global positive and negative sample pairs, the consistency of users and video content under different feature views is effectively constrained, improving the generalization ability of the model when facing complex data, realizing more efficient and accurate personalized video content recommendation, and significantly improving user experience. In view of the technical problem that in the existing recommendation video content sorting method, the optimization algorithm has limitations, resulting in poor recommendation video content sorting effect, thereby affecting the intelligent recommendation effect of video content, the present application innovatively designs a sorting optimization objective function combining user interests and business constraints, and improves the optimization algorithm through population role differentiation search strategy and dynamic inertia weight. The optimal recommended video content sorting combination is globally optimized, solving the traditional problem of limitation of optimization algorithm, improving the accuracy and personalization degree of recommendation sorting, improving the balance between user experience and business goals, realizing the optimization of recommended video content sorting, providing intelligent recommendation that meets both user needs and business constraints, and significantly improving the effect of the video content recommendation system.

[0005] The technical solutions adopted by the present application are as follows: The IPTV set-top box video content intelligent recommendation method based on deep learning provided by the present application comprises the following steps:

[0006] Step S1: data collection;

[0007] Step S2: data optimization processing;

[0008] Step S3: personalized video content recommendation;

[0009] Step S4: recommended video content optimization ranking;

[0010] Step S5: video content intelligent recommendation.

[0011] Further, in step S1, the data collection, specifically by collecting data on the IPTV set-top box terminal and the video content management platform, obtains the video content intelligent recommendation raw data; the video content intelligent recommendation raw data includes historical video content recommendation data and real-time video content recommendation data, and the historical video content recommendation data and the real-time video content recommendation data both include user viewing data, user interaction data, video content data, user portrait data and recommended environment data; the historical video content recommendation data further includes historical video content recommendation results.

[0012] Further, in step S2, the data optimization processing specifically includes the following steps:

[0013] Step S21: data cleaning, specifically data deduplication, outlier removal and null value processing on the raw data;

[0014] Step S22: data normalization processing, specifically including numerical field standardization, category field encoding and time processing;

[0015] Step S23: feature selection, specifically filtering multi-dimensional features related to personalized video content recommendation from the raw data after data cleaning and data normalization processing through correlation analysis method, to obtain a video content recommendation feature set.

[0016] Further, in step S3, the personalized video content recommendation specifically includes the following steps:

[0017] Step S31: building a personalized video content recommendation model, specifically including the following steps:

[0018] Step S311: user video preference weight calculation, specifically first calculating the probability distribution of user behavior feedback for each video content feature, then calculating the information entropy value of each video content feature through information entropy method, and finally converting the information entropy value through normalization processing to obtain the user video preference weight;

[0019] Step S312: user video interaction graph construction, used to construct multiple user video interaction graph structures by combining the interaction intensity between users and videos under each video content feature; specifically including constructing a node set, constructing a graph structure, calculating the weight of an edge and generating an adjacency matrix of the interaction graph, to obtain multiple user video interaction graph structures and their adjacency matrix set;

[0020] constructing a set of user nodes and a set of video content nodes ;

[0021] constructing a graph structure, specifically, establishing an edge for the case that a user and a video exist interaction under a video content feature, obtaining a user-video interaction graph wherein denotes a set of edges under the view, denotes an overall interaction relationship graph structure of a user and a video under the kth video content feature;

[0022] calculating the weight of the edge, specifically, for each user-video interaction graph, counting the interaction intensity of the user and the video under the feature, performing normalization processing on the interaction intensity, introducing the user-video preference weight of the corresponding video content feature for multiplication weighting, obtaining a weighted interaction value, and mapping the weighted interaction value through a Sigmoid function to obtain the weight of the edge;

[0023] generating an adjacency matrix of the interaction graph, specifically, filling the edge weight of each user-video interaction graph into the adjacency matrix according to the user-video node index corresponding position, filling 0 if there is no interaction, thereby obtaining a set of adjacency matrices;

[0024] Step S313: local interest vector generation, specifically, for each set of adjacency matrices , using a graph attention mechanism to calculate the attention coefficient between the user node and the video node, and then based on the attention coefficient, performing weighted summation on the feature vectors of the neighbor nodes to obtain a single-head local interest vector of the user under the kth video content feature view, introducing a multi-head attention mechanism to calculate the local interest vectors of multiple subspaces under the same video content feature view and perform splicing to form a comprehensive local interest vector of the video content feature, splicing the comprehensive local interest vectors of all video content feature views, and finally obtaining a set of multi-view local interest vectors of the user;

[0025] Step S314: global interest vector generation, specifically, inputting the user portrait data and the recommendation environment data into a multi-layer perception machine to perform nonlinear mapping to obtain a context query vector, using the query vector to calculate the relevance with the comprehensive local interest vectors of each video content feature view, and performing normalization through a Softmax function to obtain cross-view attention weights, and finally performing weighted summation on the interest vectors of each view according to the attention weights to obtain a global interest vector of the user;

[0026] Step S315: personalized recommendation score output, specifically, splicing the global interest vector of the user and the candidate video content vector, inputting into a fully connected layer to obtain a personalized recommendation score;

[0027] Step S32: Recommendation model training. Specifically, historical video content recommendation data is used as model training data. A multi-level contrastive loss function is used as the supervised training objective function. The model weight parameters are updated by backpropagation through the gradient descent parameter iterative update algorithm. The loss function value is gradually minimized until convergence, and the model training is completed, resulting in a trained personalized video content recommendation model.

[0028] The multi-level contrast loss function is used to simultaneously constrain the feature consistency between the user and the video under a single video content feature view and a cross-feature view. Specifically, local positive and negative samples and global positive and negative samples are constructed first, and the design of local contrast loss function and global contrast loss function is completed in sequence. Finally, the local contrast loss value and global contrast loss value are weighted and combined with the supervision result loss to obtain the multi-level contrast loss function.

[0029] ;

[0030] ;

[0031] ;

[0032] In the formula, This represents the value of the multi-level contrastive loss function. This represents the local contrast loss value. This represents the global comparison loss value. This indicates the total number of user nodes. Indicates that user u is in the first... A comprehensive local interest vector under a video content feature view. This indicates that other users unrelated to user u are in the first... A comprehensive local interest vector under a video content feature view. Represents the cosine similarity function. This represents a temperature parameter, with a range of values. , This represents the global embedding vector of the k-th video content feature view. Indicates the first Global embedding vectors of video content feature views. Represents a global embedding vector for feature views of other unrelated video content. Indicates a local negative sample. Indicates a global negative sample. The supervised loss value representing the recommendation rating is obtained through the binary cross-entropy loss function. , and respectively represent the weight coefficients of the result loss, the local contrast loss and the global contrast loss, represents the comprehensive local interest vector of the user node u under k video content feature views;

[0033] Step S33: video content personalized recommendation, specifically, taking the real-time video content recommendation data as the input data of the trained personalized video content recommendation model to obtain the recommendation result of the user video content; the recommendation result of the user video content includes a recommended video content list and a recommendation score corresponding to each video content;

[0034] Step S34: recommended video content screening, specifically, according to the recommendation result of the user video content, the video content is sorted in descending order of the recommendation score, and the first Q video contents are selected to form a video content recommendation combination to obtain a recommended video content candidate list.

[0035] Further, in step S4, the recommended video content is optimized and sorted, specifically including the following steps:

[0036] Step S41: optimization individual coding of the recommended candidate sequence, specifically, converting the recommended video content candidate list into a search individual in the optimization search space; the search individual is represented by a permutation sequence with a length of Q, where Q is the number of video content in the recommendation list;

[0037] Step S42: design a sorting optimization objective function, which is used to construct a sorting optimization objective function combining user interest and business constraints, specifically, calculate a weighted interest score based on the recommendation score and position weight of the video content, then subtract the weighted sum of three types of business penalty items from the total score in turn, and finally obtain an optimization objective function that combines user interest and business constraints; the business penalty items include an unwatched priority penalty item, a 24-hour exposure frequency control penalty item and a same type adjacent penalty item;

[0038] Step S43, recommended video content position optimization, specifically, using an improved optimization algorithm to globally optimize and sort the display position of the recommended video content candidate list in the recommendation list to obtain a video content recommendation optimization list, including the following steps:

[0039] Step S431: initialize the search population, specifically, generate Nm position vectors of search individuals by random initialization, and each position vector encoding represents a candidate video content permutation combination, and all generated search individuals together constitute an initial search population;

[0040] Step S432: calculate the individual fitness value, specifically, evaluate the fitness of each search individual in the population by the sorting optimization objective function, and calculate the search individual fitness value ;

[0041] Step S433: Calculate the search parameters, specifically, calculate the exploration coefficient and the inertia weight of the current iteration; the formula used is as follows:

[0042] ;

[0043] ;

[0044] In the formula, represents the exploration coefficient, represents the current iteration number, represents the maximum iteration number, represents the d-th dimensional position of the i-th individual in the j-th generation population, represents the average position of the current iteration population, represents the inertia weight of the current iteration, represents the maximum value of the inertia weight, represents the minimum value of the inertia weight; Step S434: Local exploration, specifically, when

[0045] , the local exploration operation is performed; the formula used is as follows:

[0046] ;

[0047] In the formula, represents the d-th dimensional position of the i-th individual in the j-th generation population, d represents the index of the dimension, represents the position of the optimal individual of the current iteration, represents the lower boundary value of the first dimension of the search space, represents the upper boundary value of the first dimension of the search space; Step S435: Global development exploration, specifically, when

[0048] , the global development operation is performed, first, the population is sorted from high to low according to the individual fitness value, and the population is divided into four categories according to the proportion, the top 40% of the individual fitness ranking is defined as the leader, the individual fitness ranking in the 40%-80% interval is defined as the explorer, the individual fitness ranking in the 80%-90% interval is defined as the follower, and the individual fitness ranking in the last 10% is defined as the laggard, a population role differentiation search strategy is adopted, and leader position updating, explorer position updating, follower position updating, and laggard position updating are performed;

[0049] ​​Step S436: obtaining the optimal position, specifically, after each iteration, the fitness values of all individuals in the current population are re-evaluated, and if the fitness of a certain individual position is better than the current global optimal individual position, the global optimal individual position is updated with the individual position;

[0050] Step S437: search iteration termination, specifically, when the search individual fitness value is higher than the fitness threshold or the maximum number of iterations is reached, the search is terminated and the global optimal individual position is obtained, and the global optimal individual position specifically refers to the video content recommendation optimization list.

[0051] Further, in step S5, the video content intelligent recommendation, specifically, based on the video content recommendation optimization list, the optimized and sorted video content is transmitted to the IPTV set-top box terminal, and the terminal displays the recommended video content according to the user interface layout strategy, realizing personalized video content intelligent recommendation and display.

[0052] The beneficial effects achieved by the above-mentioned scheme are as follows:

[0053] (1) For the technical problem that in the traditional IPTV set-top box video content recommendation method, only user preferences or business constraints are considered, resulting in a balance between user preferences and business needs that cannot be achieved in the recommended results, thus causing inaccurate recommendation results and poor user experience, the present application innovatively proposes personalized video content recommendation based on user preferences and optimized sorting of recommended video content based on business constraints, which effectively improves the accuracy of the recommended results, enhances the user experience, realizes the balance between personalization and business needs, and significantly improves the intelligent recommendation effect of IPTV video content.

[0054] (2) For the technical problem that in the existing personalized video content recommendation model, it is difficult to fully mine the complex relationship between users and video content, resulting in limited recommendation effect and inability to accurately match user interests, the present application innovatively proposes a combination method of multi-view user video interaction graph construction, double-layer attention mechanism and multi-level contrast loss function design, which can more comprehensively capture the potential connection between user behavior and video content through multi-view user video interaction graph construction, improve the personalization and accuracy of the recommendation system through the double-layer attention mechanism, which considers the user's multi-dimensional and multi-perspective interests, effectively constrain the consistency of users and video content under different feature views through the design of multi-level contrast loss function combined with local and global positive and negative sample pairs, improve the generalization ability of the model when facing complex data, realize more efficient and accurate personalized video content recommendation, and significantly improve the user experience.

[0055] (3) In view of the technical problem that the limitation of the optimization algorithm in the existing method suitable for sorting of recommended video content leads to poor sorting effect of the recommended video content, thereby affecting the intelligent recommendation effect of the video content, the scheme innovatively designs a sorting optimization objective function combining user interest and business constraint, and improves the optimization algorithm through a population role differentiation search strategy and a dynamic inertia weight, globally optimizes the optimal recommended video content sorting combination, solves the traditional problem of the limitation of the optimization algorithm, improves the accuracy and individuality of the recommendation sorting, balances the user experience and the business goal, optimizes the sorting of the recommended video content, provides intelligent recommendation meeting the user demand and satisfying the business constraint, and significantly improves the effect of the video content recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A flowchart of the IPTV set-top box video content intelligent recommendation method based on deep learning provided by the present application is shown in the figure.

[0057] Figure 2 A flowchart of step S3 is shown in the figure.

[0058] Figure 3 A flowchart of step S4 is shown in the figure.

[0059] Figure 4 A flowchart of step S31 is shown in the figure.

[0060] Figure 5 A flowchart of step S43 is shown in the figure.

[0061] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0063] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the systems or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0064] In one embodiment, referring to Figure 1 The technical scheme adopted by the present application is as follows: The IPTV set-top box video content intelligent recommendation method based on deep learning provided by the present application comprises the following steps:

[0065] Step S1: data collection, used for acquiring original data for constructing a video content intelligent recommendation system, specifically, by collecting historical data and real-time data related to video recommendation, video content intelligent recommendation original data is obtained;

[0066] Step S2: data optimization processing, used for converting the original data into optimized data suitable for a deep learning model, specifically, by data cleaning, data normalization processing and feature selection, the data quality and the correlation of model input features are improved, and finally video content intelligent recommendation optimized data is obtained;

[0067] Step S3: personalized video content recommendation, used for generating an initial personalized video content recommendation set based on a user, specifically, by constructing a personalized video content recommendation model, using historical data as training data, and combining a multi-level contrast loss function as a supervised training objective function to train the model, a trained personalized video content recommendation model is obtained, real-time video content recommendation data is input into the trained recommendation model, a recommendation result of the user video content is output, the video content is sorted from high to low according to the recommendation score, and the first Q video contents are selected to form a recommended video content candidate list;

[0068] Step S4: recommended video content optimization sorting, used for optimizing and sorting the display position of the video content in the recommended video content candidate list based on business constraints and user preferences, specifically, the recommended video content candidate list is first converted into a search individual in an optimized search space, then a sorting optimization objective function combining user interest and business constraints is designed, finally, an improved optimization algorithm is adopted to globally optimize and sort the video content display position in the recommended video content candidate list, and a video content recommendation optimization list is obtained;

[0069] Step S5: video content intelligent recommendation, used for realizing video content intelligent recommendation, specifically, based on the video content recommendation optimization list, personalized video content intelligent recommendation and display are realized.

[0070] By performing the above operation, in view of the technical problem that in the traditional IPTV set-top box video content recommendation method, only user preference or business constraints are considered, resulting in that the recommended result cannot achieve the balance between user preference and business demand, and thus the recommendation effect is not accurate and the user experience is poor, the application innovatively proposes personalized video content recommendation based on user preference and recommended video content optimization sorting based on business constraints, by simultaneously considering user preference and business constraints, the accuracy of the recommended result is effectively improved, the user experience is enhanced, the balance between personalization and business demand is achieved, and the intelligent recommendation effect of IPTV video content is significantly improved.

[0071] Embodiment two, refer to Figure 1 , this embodiment is based on the above embodiment, in step S1, the data collection is used to obtain the original data for constructing the video content intelligent recommendation system, specifically by collecting historical data and real-time data related to video recommendation on the IPTV set-top box terminal and the video content management platform, obtaining the video content intelligent recommendation original data; the video content intelligent recommendation original data includes historical video content recommendation data and real-time video content recommendation data, the historical video content recommendation data and the real-time video content recommendation data both include user watching data, user interaction data, video content data, user portrait data and recommendation environment data; the historical video content recommendation data further includes historical video content recommendation result; the user watching data includes watching video, watching time, watching start time, watching end time, watching progress, pause and play time; the user interaction data includes click behavior, sliding behavior, comment times, channel switching and search behavior; the video content data is used to describe the basic information of the video, and enhances the content understanding ability of the recommendation system; including video, video title, video category, actor information, video label, video duration, video type, video language, video play times and video play completion rate; the user portrait data includes age, region, watching history, user subscription information and social behavior; the recommendation environment data includes time period information, weather information and holiday information.

[0072] Embodiment three, refer to Figure 1 , this embodiment is based on the above embodiment, in step S2, the data optimization processing is used to convert the original data into high-quality optimization data suitable for the deep learning model, specifically by data cleaning, data normalization processing and feature selection, obtaining the video content intelligent recommendation optimization data, including the following steps:

[0073] Step S21: data cleaning, used for cleaning invalid or redundant original data, improving data quality, specifically for original data, data deduplication, outlier elimination and null value processing;

[0074] The data deduplication is used to remove duplicate raw data, ensuring the accuracy and consistency of the data, specifically identifying and removing duplicate data entries;

[0075] The outlier removal is used to identify and remove outliers in the data that do not conform to the normal behavior pattern, specifically identifying and removing outliers in the raw data based on the Z-Score method;

[0076] The null value processing is used to maintain the integrity of the data, specifically filling in missing fields in the raw data using the mean filling method;

[0077] Step S22: Data normalization processing, for converting raw data into a standard format, specifically including numerical field standardization, category field encoding and time processing;

[0078] The numerical field standardization processing is specifically through the use of the minimum-maximum normalization method to process the numerical type field, standardized to the same range;

[0079] The category field encoding is specifically through label encoding to convert category variables into numerical form;

[0080] The time processing is specifically converting the timestamp of the raw data into a standardized time format;

[0081] Step S23: Feature selection, specifically through correlation analysis method, from the raw data after data cleaning and data normalization processing, filtering the multi-dimensional features related to personalized video content recommendation, obtaining the video content recommendation feature set.

[0082] Embodiment four, see Figure 1 、 Figure 2 and Figure 4 , this embodiment is based on the above embodiment, in step S3, the personalized video content recommendation, specifically including the following steps:

[0083] Step S31: Building a personalized video content recommendation model, specifically including the following steps:

[0084] Step S311: User video preference weight calculation, for calculating the influence of each video content feature on user preference; Specifically, first calculate the probability distribution of user behavior feedback for each video content feature, then calculate the information entropy value of each video content feature through information entropy method, finally, through normalization processing, the information entropy value is converted to obtain the user video preference weight;

[0085] The video content feature is obtained by feature selection from video content data, which describes various attributes and information of video content;

[0086] The formula is as follows:

[0087] ;

[0088] ;

[0089] ;

[0090] wherein, represents the probability of the i-th user behavior feedback under the k-th video content feature, represents the statistical value of the i-th user behavior under the k-th video content feature, and m represents the total number of user behavior feedbacks, represents the information entropy value of the k-th video content feature, reflecting the uncertainty of user behavior, represents the user video preference weight of the k-th video content feature, and n represents the total number of video content features, represents the information entropy value of the k-th video content feature;

[0091] Step S312: User video interaction graph construction, for combining the interaction intensity between users and videos under each video content feature, constructing multiple user video interaction graph structures; specifically including constructing a node set, constructing a graph structure, calculating the weight of an edge, and generating an adjacency matrix of the interaction graph, obtaining multiple user video interaction graph structures and an adjacency matrix set thereof;

[0092] The node set is constructed, specifically a user node set and a video content node set ; the user node set represents the set of all users; the video content node set represents the set of all video contents;

[0093] The graph structure is constructed, specifically the existence of interaction between users and videos under video content features is established as an edge, obtaining a user video interaction graph wherein represents the edge set under the view, represents the overall interaction relationship graph structure between users and videos under the k-th video content feature;

[0094] The weight of the edge is calculated, specifically for each user video interaction graph, the interaction intensity between users and videos under the feature is counted, the interaction intensity is normalized, the corresponding user video preference weight of the video content feature is multiplied and weighted, the weighted interaction value is obtained, and the weighted interaction value is mapped through the Sigmoid function to obtain the weight of the edge;

[0095] The statistical analysis of user-video interaction intensity under this feature specifically involves determining video content features, then filtering out user video viewing and interaction records related to this feature, counting the number of related behaviors, and obtaining the interaction intensity.

[0096] The formula used is as follows:

[0097] ;

[0098] In the formula, This indicates that user u and video v are in edge weights, This represents the interaction strength between user u and video v at the k-th video content feature. This indicates normalization processing. This represents the Sigmoid activation function;

[0099] The adjacency matrix for generating the interaction graph is specifically generated by filling the adjacency matrix with the edge weights of each user video interaction graph according to the user video node index. For positions where there is no interaction, fill in 0 to obtain the adjacency matrix set;

[0100] Step S313: Local interest vector generation, used to weighted aggregate the local neighbor information of users and video nodes under a single video content feature view, thereby extracting fine-grained interests under specific feature dimensions. The view specifically refers to the user video interaction graph; specifically, it involves generating adjacency matrix sets for each user. Using a graph attention mechanism, the attention coefficient between user nodes and video nodes is calculated. Then, based on this attention coefficient, the feature vectors of neighboring nodes are weighted and summed to obtain the user's single-head local interest vector in the k-th video content feature view. A multi-head attention mechanism is then introduced to calculate and concatenate local interest vectors from multiple subspaces within the same video content feature view, forming a comprehensive local interest vector for that video content feature. Finally, the comprehensive local interest vectors from all video content feature views are concatenated to obtain the user's multi-view local interest vector set. The formula used is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] In the formula, This represents the set of neighboring video nodes for user u in the k-th video content feature view. and These represent the user u and video v input feature vectors, respectively. denotes the video t input feature vector, denotes the linear transformation weight matrix of the kth video content feature view, the hth attention head, denotes the linear transformation weight matrix of the kth video content feature view, the hth attention head, the contribution weight of the video v to the user u, denotes the inversion operation, denotes the vector concatenation operation, denotes the single-head local interest vector of the kth video content feature view, the hth attention head, and the user node u, denotes the comprehensive local interest vector of the user u under the k video content feature views, H denotes the number of attention heads in the multi-head attention mechanism, and the value range is [4, 16], denotes the concatenation operation, denotes the kth video content feature view, the 1st attention head, denotes the kth video content feature view, the Hth attention head, denotes the linear transformation weight matrix;

[0105] Step S314: Global interest vector generation, for obtaining the global interest vector of the user under the current portrait and environmental conditions, specifically, inputting the user portrait data and the recommended environmental data into the multi-layer perception machine to perform nonlinear mapping to obtain a context query vector, calculating the correlation between the query vector and the comprehensive local interest vector of each video content feature view, and obtaining the cross-view attention weight through the Softmax function normalization, and finally weighting and summing each view interest vector according to the attention weight to obtain the global interest vector of the user; the formula used is as follows:

[0106] ;

[0107] ;

[0108] In the formula, denotes the context query vector, denotes the attention projection matrix, denotes the cross-view attention weight, denotes the global interest vector, denotes the comprehensive local interest vector of the user u under the jth video content feature view;

[0109] Step S315: Personalized recommendation score output, specifically, concatenating the global interest vector of the user with the candidate video content vector, inputting into the fully connected layer to obtain the personalized recommendation score; the formula used is as follows:

[0110] ;

[0111] In the formula, This represents the weight matrix of the fully connected layer. The bias term parameter represents the weights of the fully connected layer. Represents the candidate video content vector. This represents a personalized recommendation rating, reflecting the user's level of liking for the video content;

[0112] Step S32: Recommendation model training. Specifically, historical video content recommendation data is used as model training data. A multi-level contrastive loss function is designed as the supervised training objective function. The model weight parameters are updated by backpropagation through gradient descent parameter iterative update algorithm. The loss function value is gradually minimized until convergence, and the model training is completed, resulting in a trained personalized video content recommendation model.

[0113] The multi-level contrast loss function is used to simultaneously constrain the consistency of user and video representation under a single video content feature view and a cross-feature view. Specifically, local positive and negative samples and global positive and negative samples are constructed first, and the design of local contrast loss function and global contrast loss function is completed in sequence. Finally, the local contrast loss value and global contrast loss value are weighted and combined with the supervision result loss to obtain the multi-level contrast loss function.

[0114] The local positive and negative sample pairs are specifically composed of local positive samples formed by node representations under the video content feature view that the user is interested in, and local negative samples formed by node representations under the video content feature view with low interaction correlation and interaction threshold.

[0115] The global positive and negative sample pairs are specifically composed of global representations from different video content feature views that the same user is interested in, forming global positive samples, and global representations from unrelated or low-related video content feature views serving as global negative samples; the formula used is as follows:

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] In the formula, This represents the value of the multi-level contrastive loss function. This represents the local contrast loss value. This represents the global comparison loss value. This indicates the total number of user nodes. Indicates that user u is in the first... A comprehensive local interest vector under a video content feature view. This indicates that other users unrelated to user u are in the first... A comprehensive local interest vector under a video content feature view. Represents the cosine similarity function. This represents a temperature parameter, with a range of values. , This represents the global embedding vector of the k-th video content feature view. Indicates the first Global embedding vectors of video content feature views. Represents a global embedding vector for feature views of other unrelated video content. Indicates a local negative sample. Indicates a global negative sample. The supervised loss value representing the recommendation rating is obtained through the binary cross-entropy loss function. , and These represent the weighting coefficients for outcome loss, local contrast loss, and global contrast loss, respectively.

[0121] Step S33: Personalized video content recommendation, specifically, using real-time video content recommendation data as input data for the trained personalized video content recommendation model to obtain the recommendation results of user video content; the recommendation results of user video content include a list of recommended video content and a recommendation score corresponding to each video content;

[0122] Step S34: Recommended video content filtering, specifically, based on the recommendation results of the user's video content, the video content is sorted from high to low according to the recommendation score, and the top Q video content is selected to form a video content recommendation combination to obtain a candidate list of recommended video content.

[0123] By performing the above operations, this solution addresses the technical problem in existing personalized video content recommendation models that struggle to fully uncover the complex relationships between users and video content, leading to limited recommendation performance and an inability to accurately match user interests. This solution innovatively proposes a combination of methods: constructing a multi-view user video interaction graph, a two-layer attention mechanism, and designing a multi-level contrastive loss function. The multi-view user video interaction graph construction more comprehensively captures the potential connections between user behavior and video content. The two-layer attention mechanism comprehensively considers users' multi-dimensional and multi-perspective interests, improving the personalization and accuracy of the recommendation system. The multi-level contrastive loss function, combined with local and global positive and negative sample pairs, effectively constrains the consistency between users and video content across different feature views, improving the model's generalization ability when facing complex data. This results in more efficient and accurate personalized video content recommendation, significantly enhancing the user experience.

[0124] Example 5, see Figure 1、 Figure 3 and Figure 5 The embodiment is based on the above-mentioned embodiment, and in step S4, the recommended video content is optimized and ranked, specifically including the following steps:

[0125] Step S41: encoding of the optimized individual of the recommended candidate sequence, specifically converting the recommended video content candidate list into a search individual in the optimization search space; the search individual is represented by a permutation sequence with a length of Q, where Q is the number of recommended list video contents, and each element in the sequence corresponds to a unique identifier of a candidate video content;

[0126] Step S42: design of a ranking optimization objective function, which is used to construct a ranking optimization objective function combining user interest and business constraints, specifically calculating a weighted interest score based on the recommendation score and position weight of the video content, then subtracting the weighted sum of three types of business penalty items from the total score in turn, and finally obtaining an optimization objective function that combines user interest and business constraints; the business penalty items include an unwatched priority penalty item, a 24-hour exposure frequency penalty item, and a same-type adjacent penalty item;

[0127] ;

[0128] ;

[0129] In the formula, represents the ranking optimization objective function, represents the position index of the video content in the recommended list, represents the position weight of the qth video content, represents the recommendation score of the user for the video content at the qth position, represents the unwatched priority penalty item, which gives a penalty to the video content that has been watched within the last D days, reducing its priority in the ranking, represents the penalty for the exposure frequency of the same video exceeding the threshold within 24 hours, avoiding repeated recommendations, represents the penalty when the interval between the same type of video contents in the recommended list is less than a set value, avoiding the continuous appearance of similar contents, 、 and correspond to the weight coefficients of the three types of penalty items respectively, controlling the influence degree of the penalty on the total objective, and x represents the search individual;

[0130] Step S43: optimization of the position of the recommended video content, specifically using an improved optimization algorithm to globally optimize and rank the display position of the recommended video content candidate list in the recommended list, obtaining a video content recommendation optimization list, including the following steps:

[0131] Step S431: Initialize the search population, specifically by generating Nm position vectors of search individuals through random initialization. Each position vector encodes a candidate video content permutation and combination. All the generated search individuals together constitute the initial search population.

[0132] Step S432: Calculate the individual fitness value. Specifically, this involves evaluating the fitness of each search individual in the population using the ranking optimization objective function and calculating the fitness value of the search individual. ;

[0133] Step S433: Calculate the search parameters, specifically the exploration coefficient and the inertia weight of the current iteration; the formula used is as follows:

[0134] ;

[0135] ;

[0136] In the formula, Indicates the exploration coefficient. Indicates the current iteration number. Indicates the maximum number of iterations. Indicates that the i-th individual is in the first place. The position of the d-th dimension in the population. This indicates the average position of the population in the current iteration. This represents the inertia weight of the current iteration. This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight.

[0137] Step S434: Local exploration, specifically when At that time, a local exploration operation is performed; the formula used is as follows:

[0138] ;

[0139] In the formula, Indicates that the i-th individual is in the first place. The position of the d-th dimension in the population, where d represents the index of the dimension. This indicates the position of the optimal individual in the current iteration. This represents the lower boundary value of the first dimension of the search space. This represents the upper boundary value of the first dimension of the search space;

[0140] Step S435: Global development exploration, specifically when When performing global development operations, the population is first sorted from high to low based on individual fitness values, and then divided into four roles according to the proportions. Individuals with the top 40% fitness are defined as leaders, individuals with fitness values ​​between 40% and 80% are defined as explorers, individuals with fitness values ​​between 80% and 90% are defined as followers, and individuals with fitness values ​​in the bottom 10% are defined as laggards. A population role differentiation search strategy is adopted to perform leader position updates, explorer position updates, follower position updates, and laggard position updates.

[0141] The formula used for updating the leader's position is as follows:

[0142] ;

[0143] In the formula, , and They all said Random numbers uniformly distributed within a range This indicates that the i-th leader individual is in the... The position of the d-th dimension in the population. This indicates that the i-th leader individual is in the... The position of the d-th dimension in the population;

[0144] The explorer's position is updated using the following formula:

[0145] ;

[0146] In the formula, express Random numbers uniformly distributed within a range This indicates the position of the worst-performing individual in the current iteration. This indicates that the i-th explorer individual is in the... The position of the d-th dimension in the population. This indicates that the i-th explorer individual is in the... The position of the d-th dimension in the population;

[0147] The follower position update uses the following formula:

[0148] ;

[0149] In the formula, express Random numbers uniformly distributed within a range This indicates that the i-th follower individual is in the i-th position. The position of the d-th dimension in the population. This indicates that the i-th follower individual is in the i-th position. The dth dimension position in the dth generation population;

[0150] The laggard position is updated by the following formula:

[0151]

[0152] In the formula, represents a random number uniformly distributed in the range of 0 to 1, represents the dth dimension position of the ith laggard individual in the dth generation population, represents the dth dimension position of the ith laggard individual in the dth generation population, represents the dth dimension position of the ith laggard individual in the dth generation population; Step S436: obtaining the optimal position, specifically, after each iteration, the fitness values of all individuals in the current population are re-evaluated, and if the fitness of a certain individual position is better than that of the current global optimal individual position, the global optimal individual position is updated with the individual position;

[0153] Step S437: search iteration termination, specifically, when the search individual fitness value

[0154] is higher than the fitness threshold or the maximum number of iterations is reached, the search is terminated and the global optimal individual position is obtained, and the global optimal individual position specifically refers to the video content recommendation optimization list.

[0155] By performing the above operations, in view of the technical problem that the existing optimization algorithm applicable to the recommended video content sorting method has limitations, resulting in poor recommended video content sorting effect, which affects the intelligent recommendation effect of the video content, the scheme innovatively designs a sorting optimization objective function combining user interest and business constraint, and improves the optimization algorithm through population role differentiation search strategy and dynamic inertia weight, globally optimizes the optimal recommended video content sorting combination, solves the traditional problem of the limitation of the optimization algorithm, improves the accuracy and personalization degree of the recommended sorting, balances the user experience and business objectives, realizes the optimization of the recommended video content sorting, provides intelligent recommendation that meets both user needs and business constraints, and significantly improves the effect of the video content recommendation system.

[0156] Embodiment six, refer to Figure 1 This embodiment is based on the above-mentioned embodiments, in step S5, the video content intelligent recommendation, specifically, based on the video content recommendation optimization list, the video content sorted by optimization is transmitted to the IPTV set-top box terminal, and the terminal displays the recommended video content according to the user interface layout strategy, realizing personalized video content intelligent recommendation and display.

[0157] ​​​It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0158] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0159] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning, characterized by: The method includes the following steps: Step S1: Data collection. By collecting historical and real-time data, the raw data for intelligent video content recommendation is obtained. Step S2: Data optimization processing, specifically through data cleaning, data normalization, and feature selection, to obtain optimized data for intelligent recommendation of video content; Step S3: Personalized video content recommendation, used to generate a candidate list of recommended video content based on user preferences. Specifically, a personalized video content recommendation model is constructed by designing a multi-view user video interaction graph and a two-layer attention mechanism. Historical data is used as training data, and a multi-level contrastive loss function is used as the supervised training objective function to train the model, resulting in a trained personalized video content recommendation model. Real-time recommendation data is input into the trained model, and the recommended results of user video content are output. The video content is sorted from high to low according to the recommendation score, and the top Q video content is selected to form a candidate list of recommended video content. The construction of the personalized video content recommendation model specifically includes the following steps: Step S311: Calculate the user video preference weight. Specifically, first, calculate the probability distribution of user behavior feedback for each video content feature. Then, calculate the information entropy value of each video content feature using the information entropy method. Finally, transform the information entropy value through normalization to obtain the user video preference weight. Step S312: User video interaction graph construction, which combines the interaction intensity between the user and the video under various video content features to construct the user video interaction graph structure; specifically, it includes constructing a node set, constructing a graph structure, calculating the weights of the edges and generating the adjacency matrix of the interaction graph, to obtain the user video interaction graph structure and its adjacency matrix set; The set of construction nodes specifically refers to the set of construction user nodes. and video content node set ; The construction of the graph structure specifically involves establishing an edge for each interaction between the user and the video based on video content features, thus obtaining a user-video interaction graph. ,in This represents the set of edges in this view. This represents the overall interaction relationship graph structure between the user and the video under the k-th video content feature; The calculation of edge weights involves, for each user video interaction graph, statistically analyzing the interaction intensity between the user and the video under that feature, normalizing the interaction intensity, multiplying and weighting the corresponding user video preference weights of the video content features to obtain a weighted interaction value, and then mapping the weighted interaction value through the Sigmoid function to obtain the edge weights. The adjacency matrix for generating the interaction graph is specifically generated by filling the adjacency matrix with the edge weights of each user video interaction graph according to the user video node index. For positions where there is no interaction, fill in 0 to obtain the adjacency matrix set; Step S313: Local interest vector generation, specifically for each adjacency matrix set Using graph attention mechanism, the attention coefficient between user nodes and video nodes is calculated. Then, based on the attention coefficient, the feature vectors of neighboring nodes are weighted and summed to obtain the user's single-head local interest vector under the k-th video content feature view. Multi-head attention mechanism is introduced to calculate the local interest vectors of the subspace under the same video content feature view and concatenate them to form the comprehensive local interest vector of the video content feature. The comprehensive local interest vectors of all video content feature views are concatenated to finally obtain the user's multi-view local interest vector set. Step S314: Global interest vector generation, specifically, inputting user profile data and recommendation environment data into a multilayer perceptron, performing nonlinear mapping to obtain a context query vector, using this query vector to calculate the correlation with the comprehensive local interest vectors of each video content feature view, and normalizing it through the Softmax function to obtain cross-view attention weights, and finally weighting and summing the interest vectors of each view according to the attention weights to obtain the user's global interest vector. Step S315: Personalized recommendation score output, specifically, the user's global interest vector and the candidate video content vector are concatenated and input into the fully connected layer to obtain the personalized recommendation score; The multi-level contrast loss function is used to simultaneously constrain the feature consistency between the user and the video under a single video content feature view and a cross-feature view. Specifically, local positive and negative samples and global positive and negative samples are constructed first, and the design of local contrast loss function and global contrast loss function is completed in sequence. Finally, the local contrast loss value and global contrast loss value are weighted and combined with the supervision result loss to obtain the multi-level contrast loss function. ； ； ； In the formula, This represents the value of the multi-level contrastive loss function. This represents the local contrast loss value. This represents the global comparison loss value. This indicates the total number of user nodes. Indicates that user u is in the first... A comprehensive local interest vector under a video content feature view. This indicates that other users unrelated to user u are in the first... A comprehensive local interest vector under a video content feature view. Represents the cosine similarity function. This represents a temperature parameter, with a range of values. , This represents the global embedding vector of the k-th video content feature view. Indicates the first Global embedding vectors of video content feature views. Represents a global embedding vector for feature views of other unrelated video content. Indicates a local negative sample. Indicates a global negative sample. The supervised loss value representing the recommendation rating is obtained through the binary cross-entropy loss function. , and These represent the weighting coefficients for outcome loss, local contrast loss, and global contrast loss, respectively. This represents the comprehensive local interest vector of user node u under the k-th video content feature view; Step S4: Optimize and sort recommended video content. Specifically, design a sorting optimization objective function, improve the optimization algorithm through a population role-differentiated search strategy and dynamic inertia weight, and use the improved optimization algorithm to globally optimize and sort the display positions of video content in the candidate list of recommended video content to obtain the optimized list of recommended video content. Step S5: Intelligent video content recommendation, specifically, based on the optimized video content recommendation list, to achieve personalized intelligent recommendation and display of video content.

2. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that: In step S3, the personalized video content recommendation specifically includes the following steps: Step S31: Construct a personalized video content recommendation model; Step S32: Recommendation model training. Specifically, historical video content recommendation data is used as model training data. A multi-level contrastive loss function is used as the supervised training objective function. The model weight parameters are updated by backpropagation through the gradient descent parameter iterative update algorithm. The loss function value is gradually minimized until convergence, and the model training is completed, resulting in a trained personalized video content recommendation model. Step S33: Personalized video content recommendation, specifically, using real-time video content recommendation data as input data for the trained personalized video content recommendation model to obtain the recommendation results of user video content; the recommendation results of user video content include a list of recommended video content and a recommendation score corresponding to each video content; Step S34: Recommended video content filtering, specifically, based on the recommendation results of the user's video content, the video content is sorted from high to low according to the recommendation score, and the top Q video content is selected to form a video content recommendation combination to obtain a candidate list of recommended video content.

3. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that: In step S4, the recommended video content is optimized and sorted, specifically including the following steps: Step S41: Encoding the optimized individual of the recommended candidate sequence, specifically, converting the recommended video content candidate list into search individuals in the optimized search space; the search individuals are represented by a permutation sequence of length Q, where Q is the number of video content in the recommended list; Step S42: Design a ranking optimization objective function to construct a ranking optimization objective function that combines user interests and business constraints. Specifically, it calculates a weighted interest score based on the recommendation score and position weight of video content, and then deducts the weighted sum of three types of business penalty items from this weighted interest score in sequence, finally obtaining the optimization objective function that integrates user interests and business constraints; the business penalty items include the unwatched priority penalty item, the 24-hour exposure frequency control penalty item, and the adjacent penalty item of the same type; Step S43, recommend video content position optimization, specifically, using an improved optimization algorithm to globally optimize and sort the display positions of the candidate list of recommended video content in the recommendation list, to obtain an optimized list of recommended video content.

4. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 3, characterized in that: Step S43, the optimization of the recommended video content location, specifically includes the following steps: Step S431: Initialize the search population, specifically by generating Nm position vectors of search individuals through random initialization. Each position vector encodes a candidate video content permutation and combination. All the generated search individuals together constitute the initial search population. Step S432: Calculate the individual fitness value. Specifically, this involves evaluating the fitness of each search individual in the population using the ranking optimization objective function and calculating the fitness value of the search individual. ; Step S433: Calculate the search parameters, specifically the exploration coefficient and the inertia weight of the current iteration; the formula used is as follows: ； ； In the formula, Indicates the exploration coefficient. Indicates the current iteration number. Indicates the maximum number of iterations. Indicates that the i-th individual is in the first place. The position of the d-th dimension in the population. This indicates the average position of the population in the current iteration. This represents the inertia weight of the current iteration. This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight; Step S434: Local exploration, specifically when At that time, a local exploration operation is performed; the formula used is as follows: ； In the formula, Indicates that the i-th individual is in the first place. The position of the d-th dimension in the population, where d represents the index of the dimension. This indicates the position of the optimal individual in the current iteration. This represents the lower boundary value of the first dimension of the search space. This represents the upper boundary value of the first dimension of the search space; Step S435: Global development exploration, specifically when When performing global development operations, the population is first sorted from high to low based on individual fitness values, and then divided into four roles according to the proportions. Individuals with the top 40% fitness are defined as leaders, individuals with fitness values ​​between 40% and 80% are defined as explorers, individuals with fitness values ​​between 80% and 90% are defined as followers, and individuals with fitness values ​​in the bottom 10% are defined as laggards. A population role differentiation search strategy is adopted to perform leader position updates, explorer position updates, follower position updates, and laggard position updates. Step S436: Obtain the optimal position. Specifically, after each iteration, the fitness values ​​of all individuals in the current population are re-evaluated. If the fitness of an individual's position is better than the current global optimal individual position, then the global optimal individual position is updated using that individual position. Step S437: The search iteration terminates, specifically when the fitness value of the search individual is... When the fitness threshold is exceeded or the maximum number of iterations is reached, the search is terminated and the globally optimal individual position is obtained. Specifically, the globally optimal individual position refers to the video content recommendation optimization list.

5. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that: In step S5, the intelligent recommendation of video content specifically involves transmitting the optimized and sorted video content to the IPTV set-top box terminal based on the optimized video content recommendation list. The terminal then displays the recommended video content according to the user interface layout strategy, thereby achieving personalized intelligent recommendation and display of video content.

6. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that: In step S1, the data collection specifically involves collecting data from the IPTV set-top box terminal and the video content management platform to obtain raw data for intelligent video content recommendation. The raw data for intelligent video content recommendation includes historical video content recommendation data and real-time video content recommendation data. Both the historical and real-time video content recommendation data include user viewing data, user interaction data, video content data, user profile data, and recommendation environment data. The historical video content recommendation data also includes historical video content recommendation results.

7. The method for intelligent recommendation of video content for IPTV set-top boxes based on deep learning according to claim 1, characterized in that: In step S2, the data optimization process specifically includes the following steps: Step S21: Data cleaning, specifically, deduplication, outlier removal, and null value processing of the original data; Step S22: Data normalization processing, specifically including numerical field normalization, category field encoding, and time processing; Step S23: Feature selection, specifically, using correlation analysis to screen multidimensional features related to personalized video content recommendation from the raw data after data cleaning and normalization, to obtain a video content recommendation feature set.

Images