Method and system for controlling intelligent viewing terminal equipment
Through the blockchain connection between intelligent viewing terminal devices and cloud platform, the knowledge graph and machine learning model are used to analyze user behavior and dynamically generate management and control strategies, the problems of insufficient flexibility in management and control methods and inconsistent grading standards in the existing technology are solved, and the problem of inadequate grading standards is realized. Intelligent viewing behavior control ensures content security.
Patent Information
- Application Number
- CN202411195040.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-29
AI Technical Summary
The existing intelligent viewing terminal equipment has insufficient flexibility in the management and control methods, difficult to adjust dynamically, the classification standards are not unified, the list is cumbersome, and it is impossible to effectively prevent minors from being exposed to bad content.
By establishing a secure communication connection with the cloud management and control platform, using the knowledge graph and machine learning model of the intelligent viewing terminal device to extract and group identification, generating management and control strategies and structured parameters, combined with pre-trained user portrait model Analysis, dynamically determine whether the rating behavior meets the control conditions, and generate restrictive operation instructions based on the reinforcement learning algorithm.
It realizes intelligent viewing behavior control, dynamic adjustment of control strategies, improves the flexibility and security of control, adapts to different user needs, and effectively prevents minors from being exposed to bad content.
Smart Images

Figure CN118714408B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to intelligent terminal technology, and in particular to a control method and system for intelligent viewing terminal equipment. Background Art
[0002] With the popularity of smart viewing terminal devices such as smart TVs and set-top boxes, it has become an indispensable part of daily life for users to obtain various audio-visual programs and services through these devices. However, while enjoying the convenience, some problems that need to be solved have also emerged, such as minors being addicted to TV and exposed to bad content, some users watching for too long and affecting their health, and poor protection of audio-visual content copyrights. Therefore, how to effectively control smart viewing terminal devices, guide users to use them reasonably, and ensure content security has become the focus of the industry.
[0003] Traditional intelligent viewing terminal equipment management and control methods mainly include the following:
[0004] (1) Setting viewing time limits: Users can use the device's management and control function to set viewing time limits for specific time periods. When the set time is exceeded, the device will automatically shut down or lock. However, this method lacks flexibility and cannot be dynamically adjusted according to the user's actual situation.
[0005] (2) Content rating and restriction: By rating audiovisual content and setting the content levels that different age groups can watch, when users select content that exceeds their age group, they need to enter a password or parental authorization. This method can prevent minors from accessing inappropriate content to a certain extent, but the rating standards are difficult to unify and can be easily circumvented.
[0006] (3) Blacklist and whitelist: The device presets or the user customizes blacklists and whitelists of some channels and programs. Content on the blacklist cannot be viewed, while content on the whitelist can be viewed freely. This method has obvious control effects, but the list setting and maintenance are cumbersome, and it is impossible to judge the content outside the list.
[0007] (4) Hardware lock: Install a physical lock or electronic lock on the device, unlock it with a key or password, and prevent unauthorized use of the device. This method is highly reliable, but inconvenient to use, and once the lock is damaged, the control becomes ineffective.
[0008] The above traditional control methods each have their own advantages and disadvantages, and have certain limitations in practical applications. Summary of the invention
[0009] The embodiments of the present invention provide a method and system for controlling intelligent viewing terminal equipment, which can at least solve some problems in the prior art.
[0010] According to a first aspect of the embodiments of the present invention,
[0011] Provides a management and control method for smart viewing terminal equipment, including:
[0012] By establishing a secure communication connection based on blockchain technology with the cloud-based control platform, the control task issued by the cloud-based control platform is received, and the control task includes a control rule library, a control object list, and a control strategy generation model; based on the control task, the smart viewing terminal device knowledge graph and machine learning model are used to extract features and group identify the smart viewing terminal devices, and generate control strategies and structured control parameters;
[0013] After receiving the control parameters, the intelligent viewing terminal device analyzes the user viewing behavior based on the control parameters using a pre-trained user portrait extraction model based on the attention mechanism, and determines whether the current viewing behavior meets the control conditions through deep feature matching and correlation calculation; if the control conditions are not met, it enters the next round of real-time data perception and analysis phase, and cyclically executes the steps of analyzing the user viewing behavior based on the control parameters using a pre-trained user portrait extraction model based on the attention mechanism and subsequent steps until the current system time exceeds the control time interval;
[0014] If the control conditions are met, the optimal control execution action is dynamically decided based on the control behavior pattern using a reinforcement learning algorithm, and the corresponding restrictive operation instructions are generated and sent to the smart viewing terminal device to intervene in and guide the current viewing behavior; the restrictive operation instructions include at least one of the following: restricting channel switching, restricting volume adjustment, restricting power on and off, popping up control prompt information, forcing switching to a designated channel, forcing playback of designated content, intelligently recommending appropriate content, dynamically adjusting control intensity, and increasing family reminder interactions.
[0015] According to a second aspect of the embodiments of the present invention,
[0016] Provides a management and control system for intelligent viewing terminal equipment, including:
[0017] The first unit is used to establish a secure communication connection based on blockchain technology with the cloud control platform, receive the control task issued by the cloud control platform, and the control task includes a control rule library, a control object list, and a control strategy generation model; based on the control task, use the smart viewing terminal device knowledge graph and machine learning model to extract features and group identify the smart viewing terminal devices, and generate control strategies and structured control parameters;
[0018] The second unit is used for the intelligent viewing terminal device to analyze the user viewing behavior based on the control parameters using the pre-trained user portrait extraction model based on the attention mechanism after receiving the control parameters, and to determine whether the current viewing behavior meets the control conditions through deep feature matching and correlation calculation; if the control conditions are not met, the next round of real-time data perception and analysis phase is entered, and the analysis of the user viewing behavior based on the control parameters using the pre-trained user portrait extraction model based on the attention mechanism and subsequent steps are cyclically executed until the current system time exceeds the control time interval;
[0019] The third unit is used to dynamically decide the optimal control execution action based on the control behavior pattern using a reinforcement learning algorithm if the control conditions are met, generate corresponding restrictive operation instructions and send them to the smart viewing terminal device to intervene in and guide the current viewing behavior; the restrictive operation instructions include at least one of the following: restricting channel switching, restricting volume adjustment, restricting power on and off, popping up control prompt information, forcibly switching to a designated channel, forcibly playing designated content, intelligently recommending appropriate content, dynamically adjusting the control intensity, and increasing family reminder interaction.
[0020] According to a third aspect of the embodiments of the present invention,
[0021] An electronic device is provided, comprising:
[0022] processor;
[0023] a memory for storing processor-executable instructions;
[0024] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0025] A fourth aspect of the embodiments of the present invention is:
[0026] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0027] By introducing knowledge graphs to represent the semantically enhanced features of smart viewing terminal devices, we can fully explore the multi-dimensional features of devices and users, laying the foundation for subsequent management and control; using graph neural networks to learn the hidden feature representations of control objects in feature graphs, and using clustering algorithms to achieve control grouping, we can adaptively group according to control needs; using the control rule library to constrain and optimize the generated control strategies, we improve the interpretability and enforceability of the strategies; and using smart contracts to achieve the trusted distribution of control parameters to ensure the security of control.
[0028] The control condition judgment model based on deep feature matching and correlation calculation can comprehensively consider user characteristics and viewing content, dynamically judge the satisfaction of control conditions, and realize intelligent viewing behavior control. The model introduces word embedding and attention mechanism to improve the accuracy of user portrait matching and content correlation calculation. At the same time, through weighted fusion and threshold comparison, the strictness of control can be flexibly adjusted to meet different control needs. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 A schematic diagram of a flow chart of a method for controlling a smart viewing terminal device according to an embodiment of the present invention;
[0030] Figure 2 It is a schematic diagram of the structure of the management and control system of the intelligent viewing terminal device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0031] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0032] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0033] Figure 1 FIG. 1 is a flow chart of a method for controlling a smart viewing terminal device according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0034] S101. By establishing a secure communication connection based on blockchain technology with the cloud control platform, receiving the control tasks issued by the cloud control platform, the control tasks include a control rule library, a control object list, and a control strategy generation model; based on the control tasks, using the smart viewing terminal device knowledge graph and machine learning model to extract features and group identify the smart viewing terminal devices, and generate control strategies and structured control parameters;
[0035] S102. After receiving the control parameters, the intelligent viewing terminal device analyzes the user viewing behavior based on the control parameters using the pre-trained user portrait extraction model based on the attention mechanism, and determines whether the current viewing behavior meets the control conditions through deep feature matching and correlation calculation; if the control conditions are not met, the next round of real-time data perception and analysis phase is entered, and the steps of analyzing the user viewing behavior based on the control parameters using the pre-trained user portrait extraction model based on the attention mechanism and subsequent steps are executed cyclically until the current system time exceeds the control time interval;
[0036] S103. If the control conditions are met, the optimal control execution action is dynamically decided by the reinforcement learning algorithm according to the control behavior pattern, and the corresponding restrictive operation instructions are generated and sent to the smart viewing terminal device to intervene in and guide the current viewing behavior; the restrictive operation instructions include at least one of the following: restricting channel switching, restricting volume adjustment, restricting power on and off, popping up control prompt information, forcibly switching to a designated channel, forcibly playing designated content, intelligently recommending appropriate content, dynamically adjusting the control intensity, and increasing family reminder interaction.
[0037] In an optional embodiment,
[0038] Based on the control task, the knowledge graph of smart viewing terminal devices and the machine learning model are used to extract features and group the smart viewing terminal devices, and generate control strategies and structured control parameters, including:
[0039] By using the control object list and combining the pre-built knowledge graph of smart viewing terminal devices, the multi-dimensional features of the current smart viewing terminal devices are extracted and vectorized through knowledge reasoning and semantic association analysis to generate a feature graph of the current smart viewing terminal devices, wherein the multi-dimensional features include any one or more of hardware configuration, software system, network environment, and user group features;
[0040] Constructing a graph neural network; inputting the feature graph into the graph neural network, and dividing different smart viewing terminal devices into corresponding control groups based on the feature vector output by the graph neural network through the k-means clustering algorithm or the density clustering algorithm; for the control grouping results, using clustering evaluation indicators including the silhouette coefficient and the Galiski-Harabas index to measure the cohesion and separation of the control grouping, and estimating the credibility of the clustering effect, and taking the control grouping whose credibility is higher than the preset credibility threshold as the identified control grouping;
[0041] For the identified control groups, a pre-trained control strategy generation model is used, combined with the rule templates and constraints in the control rule library, to generate an optimized control strategy for the control group. The control strategy includes control time planning, control content recommendation, control mode selection, and control intensity adjustment. The generated control strategy is converted into structured control parameters, and the control parameters are securely sent to the smart viewing terminal device through the smart contract mechanism. The control parameters include control time interval, control channel set, control program list, control user portrait, and control behavior pattern.
[0042] For example, this paper proposes to use the pre-built knowledge graph of smart viewing terminal devices, combined with the list of managed objects, to extract multi-dimensional features and vectorize them for smart viewing terminal devices. The specific implementation steps are as follows:
[0043] Construct a knowledge graph of smart viewing terminal devices. A knowledge graph is a structured semantic network consisting of three basic elements: entity, relation, and attribute. This paper crawls public data on the Internet, such as smart TV model parameters, user comments, etc., and uses natural language processing technology to extract entities, relations, and attributes to construct a knowledge graph in the field of smart viewing terminal devices. The construction process of the knowledge graph can refer to the methods in the literature.
[0044] Multi-dimensional feature extraction. Given a list of controlled objects, that is, a collection of smart viewing terminal devices that need to be controlled, we use knowledge reasoning and semantic association analysis technology to extract multi-dimensional features related to the controlled objects from the knowledge graph. The multi-dimensional features considered in this paper include:
[0045] Hardware configuration: including screen size, resolution, CPU model, memory size, storage capacity and other parameters;
[0046] Software system: including operating system version, middleware version, application list and other information;
[0047] Network environment: including access network type (such as WIFI, 4G), network bandwidth, latency and other parameters;
[0048] User group characteristics: including statistical information such as user age distribution, gender ratio, and viewing preferences.
[0049] Feature vectorization. The extracted multi-dimensional features are digitized and standardized to generate feature vectors for subsequent grouping identification and control strategy generation. Common feature vectorization methods include One-Hot encoding and Word2Vec. One-Hot encoding is suitable for discrete features, such as operating system type; Word2Vec is suitable for text features, such as user tags.
[0050] Generate a feature graph. The controlled objects and their feature vectors are constructed into a heterogeneous graph, called a feature graph. The heterogeneous graph contains two types of nodes: controlled object nodes and feature nodes, as well as edges connecting the two types of nodes. The heterogeneous graph can well integrate the structural information and semantic information of the controlled objects, providing a richer feature representation for subsequent group identification.
[0051] After generating the feature graph of the smart viewing terminal device, this paper proposes to use the Graph Neural Network (GNN) to group and identify the control objects. GNN is a deep learning model based on graph structure data, which can effectively learn the feature representation of nodes in the graph. The group identification method used in this paper is mainly divided into the following steps:
[0052] Graph neural network construction. Use graph convolutional network (GCN) to learn the hidden feature representation of nodes in the feature graph. GCN uses graph Laplacian matrix to implement convolution operation, which can aggregate the structural information and attribute information of nodes. This paper builds a neural network structure with two layers of GCN. The first layer of GCN is used to learn low-dimensional features of control object nodes and feature nodes. The second layer of GCN generates the final feature representation of control object nodes based on the first layer.
[0053] Application of clustering algorithm. According to the feature vector of the controlled object output by GCN, the controlled objects are grouped using the clustering algorithm in unsupervised learning. This paper selects the k-means clustering algorithm and the density clustering algorithm (DBSCAN) for experimental comparison. Among them, the k-means algorithm requires the number of clusters k to be specified in advance, while the DBSCAN algorithm can generate clustering results adaptively. In the implementation process, both algorithms use Euclidean distance to measure the similarity between controlled objects.
[0054] Clustering evaluation. In order to evaluate the clustering effect, this paper uses two indicators: Silhouette Coefficient and Calinski-Harabasz Index. The silhouette coefficient measures the cohesion and separation of clusters, and its value range is [-1, 1]. The larger the value, the better the clustering effect. The Calinski-Harabasz Index evaluates the clustering quality by calculating the intra-class variance and inter-class variance. The larger the value, the better the clustering effect. For clustering results that meet the preset threshold (such as the silhouette coefficient is greater than 0.5), it is considered to be a credible control grouping.
[0055] After identifying the credible control groups, this paper further uses machine learning methods to generate corresponding control strategies and control parameters for different groups. The specific implementation steps are as follows:
[0056] Construct a control strategy generation model. The generation of control strategies can be regarded as a sequence generation problem, so this paper adopts a Transformer-based sequence-to-sequence model (Seq2Seq) to implement it. First, the Seq2Seq model is pre-trained using existing control strategy samples (such as control time planning, control content recommendation, etc.) to enable it to have the basic ability to generate control strategies. Then, for the identified control group, the common features of the control objects in the group (such as user group preferences, hardware configuration level, etc.) are extracted and input into the Seq2Seq model to generate a control strategy sequence adapted to the group.
[0057] Control strategy optimization. In order to improve the feasibility and rationality of the generated control strategy, this paper introduces a control rule library to constrain and optimize the generated results. The control rule library predefines a series of rule templates and constraints, such as "control time cannot exceed 6 hours a day" and "control content must comply with laws and regulations". The control strategy generated by the Seq2Seq model is matched with the rule library, the strategies that do not meet the constraints are filtered out, and the rule templates are used to fill and improve the strategies.
[0058] Generation of control parameters. Based on the optimized control strategy, structured control parameters are extracted, including control time interval, control channel set, control program list, control user profile, and control behavior pattern. These control parameters can be directly applied to the control execution of smart viewing terminal devices.
[0059] Control and release. Through the smart contract mechanism, control parameters are securely and reliably released to the corresponding smart viewing terminal devices. Smart contracts are a trusted execution environment based on blockchain technology that can ensure the integrity and immutability of control parameters. This paper uses the Ethereum platform to implement smart contracts for control and release, including encryption, signing and verification of control parameters, to ensure the secure release and execution of control strategies.
[0060] By introducing knowledge graphs to represent the semantically enhanced features of smart viewing terminal devices, we can fully explore the multi-dimensional features of devices and users, laying the foundation for subsequent management and control; using graph neural networks to learn the hidden feature representations of control objects in feature graphs, and using clustering algorithms to achieve control grouping, we can adaptively group according to control needs; using the control rule library to constrain and optimize the generated control strategies, we improve the interpretability and enforceability of the strategies; and using smart contracts to achieve the trusted distribution of control parameters to ensure the security of control.
[0061] In an optional embodiment,
[0062] By using the control object list and combining the pre-built knowledge graph of smart viewing terminal devices, the multi-dimensional features of the current smart viewing terminal devices are extracted and vectorized through knowledge reasoning and semantic association analysis to generate a feature graph of the current smart viewing terminal devices, including:
[0063] Map the current smart viewing terminal device in the control object list to the corresponding entity node in the knowledge graph of the smart viewing terminal device; take the current smart viewing terminal device entity node as the central node, generate a multi-hop association subgraph through a random walk algorithm, and the multi-hop association subgraph includes entity nodes and relationship edges that have semantic associations with the central node; on the basis of the multi-hop association subgraph, learn the knowledge-enhanced feature representation of the current smart viewing terminal device entity node through message passing and aggregation operations in the graph neural network model;
[0064] Acquire multi-dimensional heterogeneous attribute information of the entity node of the current smart viewing terminal device, and extract structured multi-dimensional features through feature templates; splice the multi-dimensional features of the entity node of the current smart viewing terminal device with the knowledge-enhanced feature representation, and learn the importance weights of features of different dimensions through the attention mechanism to generate a comprehensive feature embedding vector of the entity node of the current smart viewing terminal device; input the comprehensive feature embedding vector of the entity node in the multi-hop association subgraph into the graph neural network model, and update the comprehensive feature embedding vector through multi-layer message passing and feature aggregation;
[0065] The comprehensive feature embedding vector of the entity node of the current smart viewing terminal device is optimized through a comparative learning model; the pairwise similarity between the optimized comprehensive feature embedding vectors of the entity node of the current smart viewing terminal device is calculated to construct a similarity matrix; the similarity matrix is feature decomposed to obtain the eigenvectors corresponding to the first m maximum eigenvalues as the low-dimensional representation of the comprehensive feature embedding vector; the low-dimensional representation of the comprehensive feature embedding vector is nonlinearly reduced by using a t-distribution neighborhood embedding algorithm, and the KL divergence between nodes in the low-dimensional space is minimized through gradient descent optimization to obtain the node coordinates on the two-dimensional plane, and generate a visualized feature map of the current smart viewing terminal device.
[0066] Exemplarily, this application uses knowledge graphs and deep learning technology to mine the features of smart viewing terminal devices from multiple dimensions, generate comprehensive feature embedding vectors, and finally construct a visual feature graph. The specific implementation steps are as follows:
[0067] First, the current smart viewing terminal device in the list of managed objects is mapped to the corresponding entity node in the pre-built knowledge graph of smart viewing terminal devices. The knowledge graph is a structured semantic network consisting of three basic elements: entity, relation, and attribute. Through the mapping operation, the managed object is associated with the entity node in the knowledge graph.
[0068] Then, taking the current smart viewing terminal device entity node as the central node, a multi-hop association subgraph is generated by a random walk algorithm. The random walk algorithm is a commonly used graph sampling method that obtains a local subgraph related to the central node by randomly walking in the graph. Specifically, starting from the central node, neighbor nodes are randomly selected with a certain probability to walk until the preset number of hops or coverage requirements are met. The generated multi-hop association subgraph contains entity nodes that have semantic associations with the central node and the relationship edges connecting them.
[0069] Based on the multi-hop association subgraph, the knowledge-enhanced feature representation of the entity nodes of the current smart viewing terminal device is learned through the Graph Neural Network (GNN) model. GNN is a deep learning model based on graph structure data, which can effectively aggregate the structural information and attribute information of the node.
[0070] This paper uses Graph Attention Network (GAT) to achieve feature learning. GAT assigns different weights to neighbor nodes through the attention mechanism, and can adaptively aggregate important neighbor information. Specifically, GAT is applied on a multi-hop association subgraph to update the feature representation of the node through multi-layer message passing and aggregation operations. The calculation formula for each layer is as follows:
[0071] ;
[0072] in, Representation Node i In the l+1 The feature vector of the layer, N i represents the set of neighbor nodes of node i, represents the attention weight between node i and its neighbor node j, Indicates l The linear transformation matrix of the layer, Represents the activation function.
[0073] Through the calculation of multi-layer GAT, the knowledge-enhanced feature representation of the entity node of the current intelligent viewing terminal device can be obtained, which is recorded as ,in Lis the number of layers of GAT.
[0074] In addition to the semantic information in the knowledge graph, smart viewing terminal devices also have heterogeneous attributes in multiple dimensions, such as hardware configuration, software version, network environment, etc. In order to make full use of this attribute information, this paper proposes a method for multi-dimensional feature extraction and fusion.
[0075] First, obtain the multi-dimensional heterogeneous attribute information of the entity node of the current smart viewing terminal device, and extract the structured multi-dimensional features through the feature template. The feature template is a rule-based feature extraction method that matches and extracts attribute values by defining a series of templates. For example, for the hardware configuration dimension, you can define templates such as "CPU model: {model}" and "memory size: {size}", and then extract the corresponding values from the attribute information.
[0076] Then, the extracted multi-dimensional features are concatenated with the knowledge-enhanced feature representation to obtain a comprehensive feature vector. In order to learn the importance weights of features of different dimensions, this paper introduces the attention mechanism. Specifically, the comprehensive feature vector is linearly transformed and SoftMax normalized to obtain the attention weights of each dimensional feature. Finally, the comprehensive feature vector is multiplied by the attention weight to obtain the weighted fusion comprehensive feature embedding vector, denoted as e center .
[0077] In order to further mine the association information between nodes in the multi-hop association subgraph, this paper again applies the graph neural network model to perform message passing and feature aggregation on the comprehensive feature embedding vectors of the nodes.
[0078] Specifically, the comprehensive feature embedding vectors of all nodes in the multi-hop association subgraph form a matrix , where N is the number of nodes in the subgraph. Then, the feature matrix is convolved using a graph convolutional network (GCN) to update the feature representation of the node. The calculation formula for each layer is as follows:
[0079] ;
[0080] in, Indicates l The node feature matrix of the layer, represents the adjacency matrix of the subgraph, Degree matrix representing the adjacency matrix of the subgraph.
[0081] Through the calculation of multi-layer GCN, an updated node feature matrix can be obtained. Each row in the updated node feature matrix corresponds to the feature vector of a node, which integrates the feature information of the node itself and the neighboring nodes.
[0082] In order to further optimize the comprehensive feature embedding vector of the physical node of the current smart viewing terminal device, this paper introduces the contrastive learning model. Contrastive learning is an unsupervised representation learning method that learns the feature representation of data by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.
[0083] Specifically, the comprehensive feature embedding vector of the current smart viewing terminal device entity node is used as the anchor point (Anchor), and positive sample nodes and negative sample nodes are sampled from the multi-hop association subgraph. Positive sample nodes are nodes that are highly correlated with the anchor point, and negative sample nodes are nodes that are irrelevant or weakly correlated with the anchor point. Then, the comprehensive feature embedding vector is optimized by minimizing the following contrast loss function:
[0084] ;
[0085] in, Represents the feature embedding vector of the positive sample node, represents the feature embedding vector of the i-th negative sample node, K is the number of negative samples, sim() represents the similarity function, represents the temperature hyperparameter.
[0086] By minimizing the contrast loss function through the gradient descent method, the optimized comprehensive feature embedding vector of the entity node of the current smart viewing terminal device can be obtained.
[0087] On the basis of the optimized comprehensive feature embedding vector, the pairwise similarities between the current smart viewing terminal device entity node and other nodes are calculated to construct the similarity matrix S.
[0088] Then, the similarity matrix is eigendecomposed to obtain the eigenvectors corresponding to the first m largest eigenvalues as the low-dimensional representation of the comprehensive feature embedding vector. Eigendecomposition is a commonly used dimensionality reduction method that can map high-dimensional data to a low-dimensional space while retaining the main features of the data.
[0089] Finally, the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm is used to embed the low-dimensional representation E of the comprehensive feature vector. m Perform nonlinear dimensionality reduction and map it to a two-dimensional plane. t-SNE is a manifold learning algorithm that preserves the local structure of the data by minimizing the KL divergence between nodes in a low-dimensional space.
[0090] Specifically, the goal of the t-SNE algorithm is to find a low-dimensional embedding space so that the distribution of nodes in this space is as similar as possible to the distribution in the high-dimensional space. The following cost function is optimized by gradient descent:
[0091] ;
[0092] Among them, P represents the similarity distribution between nodes in high-dimensional space, Q represents the similarity distribution between nodes in low-dimensional space, and Represent the similarity between node i and node j in high-dimensional space and low-dimensional space respectively.
[0093] Through t-SNE dimensionality reduction, we can get the node coordinates on the two-dimensional plane. By plotting these coordinate points on the two-dimensional plane and connecting the edges according to the node type and relationship, we can get a visual feature map of the current smart viewing terminal device. The feature map intuitively shows the association relationship and relative position between the current smart viewing terminal device and other entities, providing an intuitive reference for subsequent management and control grouping and strategy generation.
[0094] So far, this application has elaborated on the complete technical solution for feature extraction and vectorization of smart viewing terminal devices. This solution makes full use of technologies such as knowledge graphs, deep learning, and dimensionality reduction visualization to mine and integrate the features of smart viewing terminal devices from multiple dimensions, generate comprehensive feature embedding vectors and visualized feature graphs, and lay the foundation for intelligent management and control.
[0095] In an optional embodiment,
[0096] For the identified control group, the pre-trained control strategy generation model is used, combined with the rule templates and constraints in the control rule library, to generate an optimized control strategy for the control group, including:
[0097] According to the attribute feature vector of each device node in the control group, the similarity matrix between the devices is calculated by cosine similarity, and the similarity matrix is divided by applying the spectral clustering algorithm to obtain device subgroups; the attribute feature vector of each device subgroup is aggregated, the representation vector of the subgroup is learned by the attention mechanism, and matched with the rule template in the control rule library, and the candidate rule set applicable to the subgroup is screened out;
[0098] The group feature vector of the control group and the representation vector of each device subgroup are concatenated to form a control context vector, which is input into a pre-trained control time planning model to generate candidate control time plans; a feasibility analysis is performed on each candidate control time plan, and infeasible time plans are screened out through constraint satisfaction calculation and rule conflict detection to obtain a control time planning sequence; the behavior feature vector of the control group and the control time planning sequence are input into a pre-trained control content recommendation model, and a collaborative filtering algorithm based on a graph neural network is used to realize user-content interaction modeling through message passing and embedding aggregation to generate a control content recommendation list;
[0099] According to the network topology and communication mode of the fine-grained device sub-groups, a heterogeneous graph representation is constructed and input into the pre-trained control mode selection model. The aggregation representation learning of nodes and edges is realized through the graph attention network, and the graph classification algorithm is used to predict the control deployment mode selection; a control intensity evaluation index system is constructed and input into the pre-trained control intensity adjustment model to generate the control intensity adjustment configuration; the control time planning sequence, the control content recommendation list, the control deployment mode selection and the control intensity adjustment configuration are combined and optimized, and the constraint satisfaction problem is solved by the integer programming method to obtain the optimized control strategy for the control group.
[0100] Exemplarily, this application uses a pre-trained control strategy generation model, combined with the rule templates and constraints in the control rule library, through a series of machine learning and optimization algorithms, to automatically generate personalized control strategies for control groups. The specific implementation steps are as follows:
[0101] First, based on the attribute feature vector of each device node in the control group, the similarity matrix between devices is calculated by cosine similarity. Cosine similarity is a commonly used vector similarity measurement method that measures the similarity between two vectors by calculating the cosine value of the angle between them.
[0102] Then, the spectral clustering algorithm is applied to divide the similarity matrix to obtain device subgroups. Spectral clustering is a clustering method based on graph theory. It divides data into different subgroups by eigendecomposing the Laplacian matrix of the data. Specifically, the similarity matrix is converted into the adjacency matrix of an undirected weighted graph, its normalized Laplacian matrix is calculated, and the matrix is eigenvalue decomposed. The eigenvectors corresponding to the first k smallest non-zero eigenvalues are taken to form a feature matrix, which is clustered by k-means to obtain k device subgroups.
[0103] For each device subgroup, the attribute feature vectors of its internal device nodes are aggregated, and the representation vector of the subgroup is learned through the attention mechanism. The attention mechanism is a commonly used feature aggregation method that can adaptively assign weights to different features. Suppose the subgroup Gi contains ni device nodes, and its attribute feature vector matrix is Xi, then the attention mechanism is used to calculate the subgroup representation vector gi. First, Xi is nonlinearly transformed through the parameter matrices W1 and W2 and the attention distribution ai is calculated, and then ai and Xi are weighted and summed to obtain gi.
[0104] Next, the representation vector gi of each subgroup is matched with the rule template in the control rule library to filter out the candidate rule set applicable to the subgroup. The rule template defines the basic structure and parameter range of the control rule, and can be matched with the subgroup through the constraints of the attribute characteristics. The cosine similarity is used to calculate the similarity between gi and the rule template representation, and the rules with similarity greater than the preset threshold are selected as candidates.
[0105] When generating a control time plan, the group feature vector of the control group and the representation vector of each device subgroup are first concatenated to form a control context vector as the input of the control time planning model. The model uses a sequence-to-sequence (seq2seq) neural network structure to map the control context vector to the latent space through the encoder, and then generate a candidate control time sequence through the decoder.
[0106] Then, a feasibility analysis is performed on each candidate time plan. By calculating the satisfaction of the time plan with the rule constraints and detecting conflicts between different rules, infeasible time plans are screened out to obtain the final control time planning sequence. Among them, the constraint satisfaction can be calculated by measuring the degree of deviation between the time plan parameters and the rule requirements; rule conflict detection is achieved by constructing a rule dependency graph and using a graph algorithm to analyze the combination of different rules.
[0107] The control content recommendation aims to select the best content combination from the candidate control content library according to the behavioral characteristics of the control group. A collaborative filtering algorithm based on graph neural network (GNN) is used to model the relationship between users and content through the user-content interaction graph.
[0108] First, a user-content interaction graph is constructed based on the behavioral feature vectors of the control groups and the control time planning sequence. The control groups and content items are regarded as nodes of the graph, and the interactive behaviors (such as browsing, clicking, etc.) are regarded as edges of the graph to construct a heterogeneous interaction graph. Then, through the message passing and embedding aggregation mechanism of GNN, the embedded representation of user nodes and content nodes is learned to achieve user-content interaction modeling. Finally, the user's preference scores for different contents are calculated based on the similarity of node embeddings to generate a personalized control content recommendation list.
[0109] For fine-grained device sub-groupings, it is necessary to select an appropriate management and control deployment mode based on its network topology and communication mode. First, the network information and communication records of the sub-groups are constructed as a heterogeneous graph representation, including device nodes, link edges, and communication edges. Then, the heterogeneous graph is input into the pre-trained graph attention network (GAT) model, and the representation vectors of the nodes and edges of the graph are learned through the attention aggregation of nodes and edges. Finally, a graph classification algorithm such as the graph convolutional network (GCN) is used to predict the management and control deployment mode based on the embedded representation of nodes and edges, so as to achieve adaptive selection of the management and control mode.
[0110] The purpose of control intensity adjustment is to balance control effect and user experience, and dynamically adjust the strictness of control strategy. First, a multi-dimensional control intensity evaluation index system is constructed, which comprehensively considers factors such as control cost, user satisfaction, and security risk to quantify the impact of control intensity. The attribute characteristics of the control group and the control history data are input into the pre-trained control intensity adjustment model, and the nonlinear mapping relationship between control intensity and evaluation indicators is learned through a deep neural network to generate the optimal control intensity adjustment configuration.
[0111] On the basis of generating control time planning, content recommendation, deployment mode selection and intensity adjustment, the overall control strategy is further obtained through combinatorial optimization. The discrete variables of the control strategy are modeled using the integer programming method. The objective function is set to maximize the comprehensive control utility. The constraints include control rule satisfaction, resource constraints, etc. The integer programming problem is solved to obtain the optimal strategy combination. This combinatorial optimization process can balance the trade-offs between different control elements and obtain the global optimal control strategy.
[0112] Through the above steps, the technical solution realizes the adaptive strategy optimization of control grouping. By comprehensively utilizing machine learning algorithms and optimization methods, systematic modeling and solving are carried out in terms of device sub-grouping, strategy element generation, and strategy combination optimization, which can automatically generate personalized and refined control strategies.
[0113] In an optional embodiment,
[0114] The group feature vector of the control group and the representation vector of each device subgroup are concatenated to form a control context vector, which is input into a pre-trained control time planning model to generate a candidate control time plan, including:
[0115] Arrange the device subgroup representation vectors and the corresponding group feature vectors in timestamp order to form a control context sequence, and use a position-based embedding method to encode the timestamp information into the vector representation; according to a multi-layer stacked bidirectional gated recurrent unit network, take the control context sequence as input, and obtain the temporal dependency between subgroups through information transfer in both forward and backward directions; apply an attention mechanism to each layer of the bidirectional gated recurrent unit network, adjust the importance of features at different time steps by calculating the similarity weights between the control context vectors, and obtain a hidden state vector after weight aggregation;
[0116] The hidden state vector of the last layer of bidirectional gated recurrent units is subjected to feature fusion through a multi-head self-attention mechanism, where each attention head independently calculates the feature interactions of different subspaces, and the outputs of all heads are concatenated to form a comprehensive feature representation of the control decision. The output of the multi-head self-attention mechanism is added to the control context sequence using a residual connection, and a layer normalization operation is applied to obtain an enhanced control context vector representation.
[0117] The enhanced control context vector representation is input into the control time attribute decoder, and the three parallel sub-decoders in the control time attribute decoder are used to process them respectively, and the probability distribution of the control start time, duration and repetition period is output. The first activation function is applied to the probability distribution of the control start time to obtain the prediction result of the control start time; the second activation function is applied to the probability distribution of the duration and the repetition period, and multiplied by the preset maximum duration and period value to obtain the prediction result of the continuous value; the prediction results of the control start time and the continuous value are combined to construct the search space of the control time plan, and the beam search algorithm is applied for heuristic search to obtain the top N candidate control time plans with the highest probability.
[0118] Exemplarily, control time planning is a key link in the optimization of control strategies, and its goal is to automatically generate the optimal control time plan based on the group characteristics of the control group and the characteristics of the device sub-group. This application proposes a control time planning model based on a deep sequence model and an attention mechanism, which can model the temporal dependencies of the control context, mine the control rules of different time granularities, and thus generate personalized control time plans.
[0119] First, the group feature vector of the control group and the representation vector of each device subgroup are concatenated to form a control context vector as the input of the control time planning model. Specifically, the representation vector of the device subgroup and the group feature vector of the corresponding control group are arranged in timestamp order to form a control context sequence. In order to introduce time information into the feature representation, a position-based embedding method is used to encode the timestamp. Position embedding maps the timestamp to a high-dimensional space through trigonometric functions, using sine functions for even dimensions and cosine functions for odd dimensions to distinguish the features of different time positions. The position embedding vector is added element by element to the control context vector to obtain the final sequence representation.
[0120] In order to capture the temporal dependencies in the control context sequence, a multi-layer stacked bidirectional gated recurrent unit (BiGRU) network is used to model the sequence. BiGRU considers both the forward and backward information transmission of the sequence, which can more comprehensively characterize the temporal interaction pattern between subgroups. At each time step, BiGRU controls the update of the hidden state through the reset gate and the update gate. The reset gate determines the importance of the past state, and the update gate controls the degree of fusion of the current input and the past state. Through the gating mechanism, BiGRU can adaptively choose to remember and forget information of different time scales. The forward and reverse hidden states are spliced at each time step to form a complete sequence representation.
[0121] The attention mechanism is introduced in each layer of the BiGRU network. By calculating the correlation between features at different time steps, the importance of features is adaptively adjusted. Using dot product attention, the hidden state of BiGRU is first mapped to three vector spaces: query, key, and value through linear transformation. Then, the dot product of the query vector and all key vectors is calculated, and the attention weight distribution is obtained after softmax normalization. Normalization ensures the comparability of attention weights at different time steps. Finally, the attention weights are weighted and summed with the corresponding value vectors to obtain the aggregated feature representation. The attention mechanism can dynamically focus on the significant features of different time steps and extract key information for time management planning.
[0122] In the last layer of the BiGRU network, a multi-head self-attention mechanism is used to further enhance the representation capability of the control context vector. Multi-head self-attention can capture more diverse dependencies by computing feature interactions in different subspaces in parallel with multiple independent attention heads. Each attention head uses different query, key, and value transformation matrices to independently perform attention calculations in the subspace. Then, the outputs of all attention heads are concatenated, and the features captured by different heads are fused through linear transformation. Multi-head self-attention can mine the diverse interaction patterns of control context in different feature subspaces and improve the representation capability of time planning.
[0123] In order to promote the optimization and generalization of the model, residual connections and layer normalization are introduced after multi-head self-attention. The output of multi-head self-attention is added element by element to the original control context sequence to form a residual connection. Residual connections can alleviate the optimization difficulties of deep networks and promote the back propagation of gradients. Then, the results of residual connections are layer normalized. Layer normalization normalizes each feature to zero mean and unit variance by subtracting the mean from the feature and dividing it by the standard deviation. After normalization, it is adjusted by scaling and offset parameters so that the model can adaptively learn feature distribution. Layer normalization can accelerate model convergence and improve training stability.
[0124] After obtaining the enhanced control context vector representation, the final control time plan is generated through the control time attribute decoder. The decoder consists of three parallel sub-decoders, which predict the start time, duration, and repetition period of the control respectively. For the start time, the softmax function is used to map the feature representation to the time step probability distribution, and the time step with the largest probability is taken as the prediction result. For the duration and repetition period, the feature representation is first mapped to the continuous value space through linear transformation, and then multiplied by the preset maximum duration and period value to obtain the final prediction result. By decoding different time attributes in parallel, a variety of control time plans can be flexibly generated.
[0125] In order to further optimize the control time plan, the beam search algorithm is used to evaluate and screen multiple candidate plans. Beam search maintains a fixed-size set of candidate plans, expands the current optimal plan at each time step, and prunes low-quality plans. By comprehensively considering the generation probability and control utility, the top N optimal control time plans are selected as the final output. Beam search can efficiently find the global optimal solution in the search space, balancing the quality of the plan and the generation efficiency.
[0126] Through the above steps, the control time planning model can automatically generate personalized control time plans based on the group characteristics of the control group and the representation vector of the device subgroup. The model comprehensively utilizes deep sequence models, attention mechanisms, and decoding optimization techniques to model the temporal dependencies of control contexts, mine multi-granular control rules, and generate flexible and diverse combinations of control time attributes. This provides an effective technical means to achieve intelligent optimization of control strategies, which can significantly improve the timeliness and accuracy of control.
[0127] In an optional embodiment,
[0128] Based on the control parameters, the user viewing behavior is analyzed using a pre-trained user portrait extraction model based on the attention mechanism, including:
[0129] The smart viewing terminal device parses the received control parameters, extracts the control start time and control end time in the control time interval, and converts them into the timestamp format inside the device; obtains the current system time through the built-in clock circuit or network time synchronization protocol, compares the obtained current system time with the start and end timestamps of the control time interval, and determines whether the current system time is within the control time interval; if the comparison result shows that the current system time is not within the control time interval, the smart viewing terminal device enters the standby monitoring state until the control effectiveness conditions are met;
[0130] If the comparison result shows that the current system time is within the control time interval, the smart viewing terminal device triggers the control effectiveness flag, indicating that it is in the control effectiveness period, and initializes the control execution environment; during the control effectiveness period, the user viewing behavior of the smart viewing terminal device is continuously obtained, and the user viewing behavior includes channel switching, program selection, and viewing time; the smart viewing terminal device pre-processes the collected user viewing behavior and inputs it into the user portrait extraction model based on the attention mechanism, and maps the pre-processed user viewing behavior to a dense vector through the embedding layer to generate a user behavior feature vector;
[0131] The generated user behavior feature vector is encoded with multi-head self-attention, and the key patterns and preferences of user behavior are extracted by calculating the similarity weights between different features. The sequence of the extracted key patterns and preferences of user behavior is modeled through a gated recurrent unit network to determine the change trend vector representing the user preference. The key patterns and preferences of user behavior are fused with the change trend vector, and a comprehensive user behavior representation vector is obtained through residual connection and layer normalization technology. Based on the comprehensive user behavior representation vector, a multi-dimensional feature portrait of the user, including demographic attributes, interest preferences, and viewing habits, is predicted.
[0132] For example, under the guidance of control parameters, the intelligent viewing terminal device needs to analyze the user's viewing behavior, extract the user's multi-dimensional feature portrait, and provide data support for personalized content recommendation and advertising. This application proposes a user portrait extraction model based on the attention mechanism, which can mine key patterns and preferences from the user's viewing behavior and predict the user's multi-dimensional features such as demographic attributes, interests and hobbies, and viewing habits.
[0133] The smart viewing terminal device first parses the received control parameters, extracts the start and end time of the control time interval, and converts it into the timestamp format inside the device. The current system time is obtained through the built-in clock circuit or network time synchronization protocol, and the current time is compared with the control time interval. If the current time is not within the control time interval, the smart viewing terminal device enters the standby monitoring state and regularly checks the control effectiveness conditions. When the control effectiveness conditions are met, that is, the current time enters the control time interval, the smart viewing terminal device triggers the control effectiveness flag, indicating that the control effectiveness period has begun, and initializes the control execution environment.
[0134] During the effective period of control, the smart viewing terminal device continuously obtains the user's viewing behavior data. Viewing behavior includes information such as channel switching, program selection, and viewing time. The collected user viewing behavior data is preprocessed, such as data cleaning, feature extraction, and normalization. The preprocessed user viewing behavior data is represented in the form of a fixed-length sequence, and each time step corresponds to a viewing behavior record.
[0135] The preprocessed user viewing behavior sequence is input into the embedding layer, and each viewing behavior is mapped to a dense vector representation through a table lookup operation. The embedding layer can learn the semantic similarity between viewing behaviors and convert discrete behavior IDs into a continuous feature space. The dimension of the embedding vector is tuned through methods such as cross-validation to balance the feature representation ability and computational complexity. The output of the embedding layer is a sequence of user behavior feature vectors, which serves as the input of the subsequent attention mechanism.
[0136] Multi-head self-attention encoding is performed on the sequence of user behavior feature vectors, and the key patterns and preferences of user behavior are extracted by calculating the correlation between different features. The multi-head self-attention mechanism includes multiple parallel attention heads, each of which independently calculates attention weights and feature aggregation. For each attention head, the sequence of user behavior feature vectors is mapped to the query vector, key vector, and value vector through linear transformation. Then, the dot product similarity between the query vector and all key vectors is calculated, and the normalized attention weight is obtained by applying the softmax function. The attention weight is weighted and summed with the corresponding value vector to obtain the feature aggregation result of the current head. Finally, the outputs of all attention heads are spliced, and the output vector sequence of multi-head self-attention encoding is obtained through linear transformation.
[0137] The vector sequence of key patterns and preferences of user behaviors obtained by multi-head self-attention encoding is input into the gated recurrent unit (GRU) network for sequence modeling. The GRU network adaptively updates the hidden state through the gating mechanism, which can capture the long-term dependency and dynamic changes of user preferences. At each time step, the GRU controls the flow and update of information through the reset gate and update gate according to the current input and the past hidden state. The reset gate determines the importance of the past state, and the update gate controls the degree of fusion of the current input and the past state. The hidden state of the GRU is extracted at the last time step of the sequence as a vector representing the trend of user preference changes.
[0138] The key patterns and preference vectors of user behavior are concatenated with the change trend vector obtained by GRU to form a preliminary user behavior representation. In order to further enhance the representation ability of features, residual connection and layer normalization techniques are introduced. The preliminary user behavior representation is added element by element to the original user behavior feature vector sequence to form a residual connection. Residual connection can alleviate the optimization problem of deep networks and promote the back propagation of gradients. Then, the result of residual connection is layer normalized. Layer normalization normalizes the feature by subtracting the mean and dividing it by the standard deviation, so that each feature has zero mean and unit variance. After normalization, it is adjusted by scaling and offset parameters so that the model can adaptively learn the feature distribution. Layer normalization can accelerate model convergence and improve training stability. The final vector is a comprehensive user behavior representation that integrates the key patterns, change trends and original behavior features of user preferences.
[0139] Based on the comprehensive user behavior representation vector, the user's multi-dimensional feature portrait is predicted through the fully connected layer and activation function. Different prediction methods are used for different types of features. For demographic attributes such as gender, age group, etc., the softmax classifier is used to map the user behavior representation to the corresponding category probability distribution, and the category with the highest probability is selected as the prediction result. For continuous value features such as interest preferences and viewing habits, linear regression is used to map the user behavior representation to the predicted value of the feature. The predicted value is compressed to between 0 and 1 through the sigmoid function, indicating the degree of normalization of the feature. The final user portrait includes multiple dimensions such as demographic attributes, interest preferences, viewing habits, etc., providing rich user features for subsequent personalized recommendations and advertising.
[0140] Through the above steps, the user portrait extraction model based on the attention mechanism can automatically learn the user's key preferences and behavior patterns from the user's viewing behavior, and predict the multi-dimensional user feature portrait. This model fully utilizes the advantages of the attention mechanism in extracting key information, and mines the long-term dependence and dynamic change trend of user behavior through multi-head self-attention encoding and gated recurrent unit network. At the same time, the introduction of residual connection and layer normalization technology enhances the model's feature representation ability and training stability.
[0141] In an optional embodiment,
[0142] Through deep feature matching and correlation calculation, determine whether the current viewing behavior meets the control conditions, including:
[0143] The control user portrait in the control parameters is converted into a control user portrait vector, wherein the control user portrait is defined in the form of a structured attribute-value pair, representing the characteristics of the target user group to be controlled, and the control user portrait vector is obtained by semantically mapping the attributes and values in the control user portrait using a word embedding model; the similarity score between the real-time user portrait vector and the control user portrait vector is calculated using a deep feature matching algorithm, wherein the real-time user portrait vector is obtained by vectorizing the multi-dimensional feature portrait of the user, and the deep feature matching algorithm uses cosine similarity as a metric function for feature matching, and uses the cosine value of the angle between the real-time user portrait vector and the control user portrait vector as a similarity score;
[0144] Perform one-hot encoding on the current channel ID and program ID to obtain a Boolean channel feature vector and program feature vector. The controlled channel set and the controlled program list are also represented as a one-hot encoded feature matrix. The channel attention weight distribution and the program attention weight distribution are calculated based on the attention mechanism. The channel attention weight distribution is multiplied element by element with the channel feature vector and summed to obtain the channel matching score. The program attention weight distribution is multiplied element by element with the program feature vector and summed to obtain the program matching score. According to the channel matching score and the program matching score, a comprehensive content relevance score is calculated by a weighted summation fusion method.
[0145] According to the similarity score between the real-time user portrait vector and the control user portrait vector and the content relevance score, the comprehensive control condition satisfaction is calculated through a weighted fusion function, and the weighted fusion function controls the contribution ratio of user portrait similarity and content relevance to the control condition through a weight coefficient; the calculated control condition satisfaction is compared with a preset threshold, and when the satisfaction exceeds the preset threshold, it is determined that the current viewing behavior meets the control condition, and corresponding control measures are taken; otherwise, it is determined that the current viewing behavior does not trigger the control condition, and no control operation is performed.
[0146] Exemplarily, after obtaining the real-time portrait of the user, the intelligent viewing terminal device needs to determine whether the current user's viewing behavior meets the control conditions defined in the control parameters. This application proposes a control condition judgment model based on deep feature matching and correlation calculation, which calculates the satisfaction of the control conditions by comprehensively considering the similarity of user portraits and the correlation of viewing content, and makes control decisions based on preset thresholds.
[0147] First, the control user portrait in the control parameters is converted into a vectorized representation. The control user portrait is defined in the form of a structured attribute-value pair, which represents the characteristics of the target user group to be controlled. For each attribute-value pair, a pre-trained word embedding model, such as Word2Vec or GloVe, is used to map the attributes and values into semantic vectors. By concatenating or averaging the attribute vector and the value vector, an embedding vector representing a single attribute-value pair is obtained. Finally, the embedding vectors of all attribute-value pairs are weighted averaged or concatenated to obtain the control user portrait vector. The control user portrait vector is represented in the form of a dense vector of fixed length, capturing the semantic characteristics of the control user group.
[0148] For example, assuming that the control user portrait is {"gender": "male", "age group": "youth", "interests": "sports events"}, the word embedding model can convert it into a 300-dimensional control user portrait vector, such as [0.2, 0.1, ..., 0.5].
[0149] Using the deep feature matching algorithm, the similarity between the real-time user portrait vector and the controlled user portrait vector is calculated. The real-time user portrait vector is obtained by vectorizing the multi-dimensional feature portrait of the user, and has the same dimension as the controlled user portrait vector. Cosine similarity is used as the metric function for feature matching, and the cosine value of the angle between the two vectors is calculated as the similarity score. The value range of cosine similarity is -1 to 1, and the larger the value, the more similar the two vectors are. By setting the similarity threshold, it can be determined whether the real-time user portrait matches the controlled user portrait.
[0150] For example, assuming that the real-time user portrait vector is [0.3, 0.2, ..., 0.6], and the cosine similarity is calculated with the above-mentioned control user portrait vector, the similarity score is 0.8, indicating that the two user portraits have a high degree of match.
[0151] For the channels and programs that the current user is watching, one-hot encoding is performed to obtain Boolean channel feature vectors and program feature vectors. One-hot encoding maps channels and programs into fixed-length binary vectors. Each dimension corresponds to a unique channel or program, and the value is 0 or 1, indicating whether the channel or program is currently being watched. At the same time, the control channel set and control program list in the control parameters are also represented as a one-hot encoded feature matrix.
[0152] Next, the channel attention weight distribution and program attention weight distribution are calculated based on the attention mechanism. For the channel feature vector, the attention mechanism is used to calculate its correlation weight with each channel in the controlled channel set. Specifically, the channel feature vector is dot-producted with the controlled channel feature matrix to obtain the attention score vector. The softmax function is applied to normalize the attention score vector to obtain the channel attention weight distribution, which indicates the degree of correlation between the currently viewed channel and each channel in the controlled channel set. Similarly, attention is calculated on the program feature vector and the controlled program feature matrix to obtain the program attention weight distribution.
[0153] Then, the weighted summation method is used to multiply the channel attention weight distribution and the channel feature vector element by element and sum them up to get the channel matching score. Similarly, the program attention weight distribution and the program feature vector are multiplied element by element and summed up to get the program matching score. The channel matching and program matching values range from 0 to 1. The larger the value, the higher the correlation between the current viewing content and the controlled content.
[0154] Finally, the channel matching score and the program matching score are combined through a weighted sum fusion method to obtain a comprehensive content relevance score. The fusion weight can be adjusted according to the importance of the channel and program to the control condition, for example, giving the channel matching a higher weight.
[0155] For example, suppose the current user is watching channel A and program B, the channel feature vector is [0, 1, 0, ..., 0], and the program feature vector is [0, 0, 1, ..., 0]. The set of controlled channels is {A, C, D}, and the list of controlled programs is {B, E, F}. Through the attention mechanism calculation, the channel attention weight distribution is [0.6, 0.3, 0.1], and the program attention weight distribution is [0.8, 0.1, 0.1]. The weighted summation results in a channel matching degree of 0.6 and a program matching degree of 0.8. Assuming the channel weight is 0.6 and the program weight is 0.4, the comprehensive content relevance score is 0.6*0.6 + 0.4*0.8 = 0.68.
[0156] Based on the similarity score and content relevance score between the real-time user portrait vector and the controlled user portrait vector, the comprehensive control condition satisfaction is calculated through a weighted fusion function. The weighted fusion function controls the contribution ratio of user portrait similarity and content relevance to the control condition through the weight coefficient. For example, a higher weight can be given to user portrait similarity. The control condition satisfaction ranges from 0 to 1. The larger the value, the more likely the current viewing behavior is to trigger the control condition.
[0157] The calculated control condition satisfaction is compared with the preset threshold. When the satisfaction exceeds the preset threshold, it is determined that the current viewing behavior meets the control condition, and the smart viewing terminal device needs to take corresponding control measures, such as blocking programs, playing prompt information, etc. Otherwise, it is determined that the current viewing behavior does not trigger the control condition, and the smart viewing terminal device continues to play the program normally without performing the control operation. The preset threshold can be adjusted according to the strictness of the control. The higher the threshold, the stricter the control condition judgment.
[0158] For example, assuming that the user portrait similarity is 0.8, the content relevance score is 0.68, the similarity weight is 0.7, and the content weight is 0.3, then the control condition satisfaction is 0.7*0.8 + 0.3*0.68 = 0.764. Assuming the preset threshold is 0.75, the current viewing behavior meets the control condition and the corresponding control measures need to be implemented.
[0159] Through the above steps, the control condition judgment model based on deep feature matching and correlation calculation can comprehensively consider user characteristics and viewing content, dynamically judge the satisfaction of control conditions, and realize intelligent viewing behavior control. The model introduces word embedding and attention mechanism to improve the accuracy of user portrait matching and content correlation calculation. At the same time, through weighted fusion and threshold comparison, the strictness of control can be flexibly adjusted to meet different control needs.
[0160] Figure 2 FIG. 1 is a schematic diagram of a structure of a management and control system for an intelligent viewing terminal device according to an embodiment of the present invention. Figure 2 As shown, the system comprises:
[0161] The first unit is used to establish a secure communication connection based on blockchain technology with the cloud control platform, receive the control task issued by the cloud control platform, and the control task includes a control rule library, a control object list, and a control strategy generation model; based on the control task, use the smart viewing terminal device knowledge graph and machine learning model to extract features and group identify the smart viewing terminal devices, and generate control strategies and structured control parameters;
[0162] The second unit is used for the intelligent viewing terminal device to analyze the user viewing behavior based on the control parameters using the pre-trained user portrait extraction model based on the attention mechanism after receiving the control parameters, and to determine whether the current viewing behavior meets the control conditions through deep feature matching and correlation calculation; if the control conditions are not met, the next round of real-time data perception and analysis phase is entered, and the analysis of the user viewing behavior based on the control parameters using the pre-trained user portrait extraction model based on the attention mechanism and subsequent steps are cyclically executed until the current system time exceeds the control time interval;
[0163] The third unit is used to dynamically decide the optimal control execution action based on the control behavior pattern using a reinforcement learning algorithm if the control conditions are met, generate corresponding restrictive operation instructions and send them to the smart viewing terminal device to intervene in and guide the current viewing behavior; the restrictive operation instructions include at least one of the following: restricting channel switching, restricting volume adjustment, restricting power on and off, popping up control prompt information, forcibly switching to a designated channel, forcibly playing designated content, intelligently recommending appropriate content, dynamically adjusting the control intensity, and increasing family reminder interaction.
[0164] According to a third aspect of the embodiments of the present invention,
[0165] An electronic device is provided, comprising:
[0166] processor;
[0167] a memory for storing processor-executable instructions;
[0168] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0169] A fourth aspect of the embodiments of the present invention is:
[0170] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0171] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.
[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for controlling a smart viewing terminal device, characterized in that: include: By establishing a secure communication connection based on blockchain technology with the cloud-based control platform, the control task issued by the cloud-based control platform is received, and the control task includes a control rule library, a control object list, and a control strategy generation model; based on the control task, the smart viewing terminal device knowledge graph and machine learning model are used to extract features and group identify the smart viewing terminal devices, and generate control strategies and structured control parameters; After receiving the control parameters, the intelligent viewing terminal device analyzes the user's viewing behavior based on the control parameters using a pre-trained user portrait extraction model based on the attention mechanism, and determines whether the current viewing behavior meets the control conditions through deep feature matching and correlation calculation; If the control conditions are not met, the next round of real-time data perception and analysis phase is entered, and the user viewing behavior is analyzed based on the control parameters using a pre-trained user portrait extraction model based on the attention mechanism and subsequent steps are executed cyclically until the current system time exceeds the control time interval; If the control conditions are met, the optimal control execution action is dynamically decided by using the reinforcement learning algorithm according to the control behavior mode, and the corresponding restrictive operation instructions are generated and sent to the smart viewing terminal device to intervene and guide the current viewing behavior; the restrictive operation instructions include at least one of the following: restricting channel switching, restricting volume adjustment, restricting power on and off, popping up control prompt information, forcibly switching to a specified channel, forcibly playing specified content, intelligently recommending appropriate content, dynamically adjusting the control intensity, and increasing family reminder interaction; Based on the control task, the knowledge graph of smart viewing terminal devices and the machine learning model are used to extract features and group the smart viewing terminal devices, and generate control strategies and structured control parameters, including: By using the control object list and combining the pre-built knowledge graph of smart viewing terminal devices, the multi-dimensional features of the current smart viewing terminal devices are extracted and vectorized through knowledge reasoning and semantic association analysis to generate a feature graph of the current smart viewing terminal devices, wherein the multi-dimensional features include any one or more of hardware configuration, software system, network environment, and user group features; Constructing a graph neural network; inputting the feature graph into the graph neural network, and dividing different smart viewing terminal devices into corresponding control groups based on the feature vector output by the graph neural network through the k-means clustering algorithm or the density clustering algorithm; for the control grouping results, using clustering evaluation indicators including the silhouette coefficient and the Galiski-Harabas index to measure the cohesion and separation of the control grouping, and estimating the credibility of the clustering effect, and taking the control grouping whose credibility is higher than the preset credibility threshold as the identified control grouping; For the identified control group, the pre-trained control strategy generation model is used, combined with the rule templates and constraints in the control rule library, to generate an optimized control strategy for the control group. The control strategy includes control time planning, control content recommendation, control mode selection, and control intensity adjustment. The generated control strategy is converted into structured control parameters, and the control parameters are securely sent to the smart viewing terminal device through the smart contract mechanism. The control parameters include control time interval, control channel set, control program list, control user portrait, and control behavior pattern. By using the control object list and combining the pre-built knowledge graph of smart viewing terminal devices, the multi-dimensional features of the current smart viewing terminal devices are extracted and vectorized through knowledge reasoning and semantic association analysis to generate a feature graph of the current smart viewing terminal devices, including: Map the current smart viewing terminal device in the control object list to the corresponding entity node in the knowledge graph of the smart viewing terminal device; take the current smart viewing terminal device entity node as the central node, generate a multi-hop association subgraph through a random walk algorithm, and the multi-hop association subgraph includes entity nodes and relationship edges that have semantic associations with the central node; on the basis of the multi-hop association subgraph, learn the knowledge-enhanced feature representation of the current smart viewing terminal device entity node through message passing and aggregation operations in the graph neural network model; Acquire multi-dimensional heterogeneous attribute information of the entity node of the current smart viewing terminal device, and extract structured multi-dimensional features through feature templates; splice the multi-dimensional features of the entity node of the current smart viewing terminal device with the knowledge-enhanced feature representation, and learn the importance weights of features of different dimensions through the attention mechanism to generate a comprehensive feature embedding vector of the entity node of the current smart viewing terminal device; input the comprehensive feature embedding vector of the entity node in the multi-hop association subgraph into the graph neural network model, and update the comprehensive feature embedding vector through multi-layer message passing and feature aggregation; The comprehensive feature embedding vector of the entity node of the current smart viewing terminal device is optimized through a comparative learning model; the pairwise similarity between the optimized comprehensive feature embedding vectors of the entity node of the current smart viewing terminal device is calculated to construct a similarity matrix; the similarity matrix is feature decomposed to obtain the eigenvectors corresponding to the first m maximum eigenvalues as the low-dimensional representation of the comprehensive feature embedding vector; the low-dimensional representation of the comprehensive feature embedding vector is nonlinearly reduced by using a t-distribution neighborhood embedding algorithm, and the KL divergence between nodes in the low-dimensional space is minimized through gradient descent optimization to obtain the node coordinates on the two-dimensional plane, and generate a visualized feature map of the current smart viewing terminal device.
2. The method according to claim 1, characterized in that For the identified control group, the pre-trained control strategy generation model is used, combined with the rule templates and constraints in the control rule library, to generate an optimized control strategy for the control group, including: According to the attribute feature vector of each device node in the control group, the similarity matrix between the devices is calculated by cosine similarity, and the similarity matrix is divided by applying the spectral clustering algorithm to obtain device subgroups; the attribute feature vector of each device subgroup is aggregated, the representation vector of the subgroup is learned by the attention mechanism, and matched with the rule template in the control rule library, and the candidate rule set applicable to the subgroup is screened out; The group feature vector of the control group and the representation vector of each device subgroup are concatenated to form a control context vector, which is input into a pre-trained control time planning model to generate candidate control time plans; a feasibility analysis is performed on each candidate control time plan, and infeasible time plans are screened out through constraint satisfaction calculation and rule conflict detection to obtain a control time planning sequence; the behavior feature vector of the control group and the control time planning sequence are input into a pre-trained control content recommendation model, and a collaborative filtering algorithm based on a graph neural network is used to realize user-content interaction modeling through message passing and embedding aggregation to generate a control content recommendation list; According to the network topology and communication mode of fine-grained device sub-groups, a heterogeneous graph representation is constructed and input into the pre-trained control mode selection model. The aggregate representation learning of nodes and edges is realized through the graph attention network, and the graph classification algorithm is used to predict the control deployment mode selection; a control intensity evaluation index system is constructed and input into the pre-trained control intensity adjustment model to generate the control intensity adjustment configuration; the control time planning sequence, the control content recommendation list, the control deployment mode selection and the control intensity adjustment configuration are combined and optimized, and the constraint satisfaction problem is solved by the integer programming method to obtain the optimized control strategy for the control group.
3. The method according to claim 2, characterized in that The group feature vector of the control group and the representation vector of each device subgroup are concatenated to form a control context vector, which is input into a pre-trained control time planning model to generate a candidate control time plan, including: Arrange the device subgroup representation vectors and the corresponding group feature vectors in timestamp order to form a control context sequence, and use a position-based embedding method to encode the timestamp information into the vector representation; according to a multi-layer stacked bidirectional gated recurrent unit network, take the control context sequence as input, and obtain the temporal dependency between subgroups through information transfer in both forward and backward directions; apply an attention mechanism to each layer of the bidirectional gated recurrent unit network, adjust the importance of features at different time steps by calculating the similarity weights between the control context vectors, and obtain a hidden state vector after weight aggregation; The hidden state vector of the last layer of bidirectional gated recurrent units is subjected to feature fusion through a multi-head self-attention mechanism, where each attention head independently calculates the feature interactions of different subspaces, and the outputs of all heads are concatenated to form a comprehensive feature representation of the control decision. The output of the multi-head self-attention mechanism is added to the control context sequence using a residual connection, and a layer normalization operation is applied to obtain an enhanced control context vector representation. The enhanced control context vector representation is input into the control time attribute decoder, and the three parallel sub-decoders in the control time attribute decoder are used to process them respectively, and the probability distribution of the control start time, duration and repetition period is output. The first activation function is applied to the probability distribution of the control start time to obtain the prediction result of the control start time; the second activation function is applied to the probability distribution of the duration and the repetition period, and multiplied by the preset maximum duration and period value to obtain the prediction result of the continuous value; the prediction results of the control start time and the continuous value are combined to construct the search space of the control time plan, and the beam search algorithm is applied for heuristic search to obtain the top N candidate control time plans with the highest probability.
4. The method according to claim 1, characterized in that Based on the control parameters, the user viewing behavior is analyzed using a pre-trained user portrait extraction model based on the attention mechanism, including: The smart viewing terminal device parses the received control parameters, extracts the control start time and control end time in the control time interval, and converts them into the timestamp format inside the device; obtains the current system time through the built-in clock circuit or network time synchronization protocol, compares the obtained current system time with the start and end timestamps of the control time interval, and determines whether the current system time is within the control time interval; if the comparison result shows that the current system time is not within the control time interval, the smart viewing terminal device enters the standby monitoring state until the control effectiveness conditions are met; If the comparison result shows that the current system time is within the control time interval, the smart viewing terminal device triggers the control effectiveness flag, indicating that it is in the control effectiveness period, and initializes the control execution environment; during the control effectiveness period, the user viewing behavior of the smart viewing terminal device is continuously obtained, and the user viewing behavior includes channel switching, program selection, and viewing time; the smart viewing terminal device pre-processes the collected user viewing behavior and inputs it into the user portrait extraction model based on the attention mechanism, and maps the pre-processed user viewing behavior to a dense vector through the embedding layer to generate a user behavior feature vector; The generated user behavior feature vector is encoded with multi-head self-attention, and the key patterns and preferences of user behavior are extracted by calculating the similarity weights between different features. The sequence of the extracted key patterns and preferences of user behavior is modeled through a gated recurrent unit network to determine the change trend vector representing the user preference. The key patterns and preferences of user behavior are fused with the change trend vector, and a comprehensive user behavior representation vector is obtained through residual connection and layer normalization technology. Based on the comprehensive user behavior representation vector, a multi-dimensional feature portrait of the user, including demographic attributes, interest preferences, and viewing habits, is predicted.
5. The method according to claim 4, characterized in that Through deep feature matching and correlation calculation, determine whether the current viewing behavior meets the control conditions, including: The control user portrait in the control parameters is converted into a control user portrait vector, wherein the control user portrait is defined in the form of a structured attribute-value pair, representing the characteristics of the target user group to be controlled, and the control user portrait vector is obtained by semantically mapping the attributes and values in the control user portrait using a word embedding model; the similarity score between the real-time user portrait vector and the control user portrait vector is calculated using a deep feature matching algorithm, wherein the real-time user portrait vector is obtained by vectorizing the multi-dimensional feature portrait of the user, and the deep feature matching algorithm uses cosine similarity as a metric function for feature matching, and uses the cosine value of the angle between the real-time user portrait vector and the control user portrait vector as a similarity score; Perform one-hot encoding on the current channel ID and program ID to obtain a Boolean channel feature vector and program feature vector. The controlled channel set and the controlled program list are also represented as a one-hot encoded feature matrix. The channel attention weight distribution and the program attention weight distribution are calculated based on the attention mechanism. The channel attention weight distribution is multiplied element by element with the channel feature vector and summed to obtain the channel matching score. The program attention weight distribution is multiplied element by element with the program feature vector and summed to obtain the program matching score. According to the channel matching score and the program matching score, a comprehensive content relevance score is calculated by a weighted summation fusion method. According to the similarity score between the real-time user portrait vector and the control user portrait vector and the content relevance score, the comprehensive control condition satisfaction is calculated through a weighted fusion function, and the weighted fusion function controls the contribution ratio of user portrait similarity and content relevance to the control condition through a weight coefficient; the calculated control condition satisfaction is compared with a preset threshold, and when the satisfaction exceeds the preset threshold, it is determined that the current viewing behavior meets the control condition, and corresponding control measures are taken; otherwise, it is determined that the current viewing behavior does not trigger the control condition, and no control operation is performed.
6. A management and control system for intelligent viewing terminal equipment, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to establish a secure communication connection based on blockchain technology with the cloud control platform, receive the control task issued by the cloud control platform, and the control task includes a control rule library, a control object list, and a control strategy generation model; based on the control task, use the smart viewing terminal device knowledge graph and machine learning model to extract features and group identify the smart viewing terminal devices, and generate control strategies and structured control parameters; The second unit is used for the intelligent viewing terminal device to analyze the user's viewing behavior based on the control parameters by using a pre-trained user portrait extraction model based on the attention mechanism after receiving the control parameters, and to determine whether the current viewing behavior meets the control conditions through deep feature matching and correlation calculation; If the control conditions are not met, the next round of real-time data perception and analysis phase is entered, and the user viewing behavior is analyzed based on the control parameters using a pre-trained user portrait extraction model based on the attention mechanism and subsequent steps are executed cyclically until the current system time exceeds the control time interval; The third unit is used to dynamically decide the optimal control execution action based on the control behavior pattern using a reinforcement learning algorithm if the control conditions are met, generate corresponding restrictive operation instructions and send them to the smart viewing terminal device to intervene in and guide the current viewing behavior; the restrictive operation instructions include at least one of the following: restricting channel switching, restricting volume adjustment, restricting power on and off, popping up control prompt information, forcibly switching to a designated channel, forcibly playing designated content, intelligently recommending appropriate content, dynamically adjusting the control intensity, and increasing family reminder interaction.
7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.
8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.
