Non-intelligent set top box double-management operation method and system
Illegal behaviors are identified through end-cloud collaborative architecture and deep learning, combined with user portrait and knowledge graph technology, the problem of single non-intelligent set-top box functions and security risks is solved, and the source of illegal behaviors and personalized content recommendations are realized, improving user experience and information Safety.
Patent Information
- Application Number
- CN202411400089.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Due to the limitations of hardware configuration and software functions, non-intelligent set-top boxes cannot flexibly expand new applications and services. The content that users can choose is a single source, and the viewing experience needs to be improved. At the same time, there is a risk of illegally tampering with firmware and installing third-party software privately, resulting in users. Feeling susceptible to adverse information.
The terminal cloud collaborative architecture is used to build an illegal set-top box recognition model, identify illegal behaviors through convolutional neural networks and reinforcement learning algorithms, generate punishment action sequences, and automatically trigger them through smart contracts; combine user portrait technology and factor decomposition machines to predict user preferences and make differences To upgrade and content recommendation; build a whole-region knowledge graph and generate a personalized recommendation list through collaborative filtering algorithm.
It has achieved accurate identification and traceability of illegal behaviors, improved user experience, provided personalized content recommendations, enhanced information security and governance effects, and significantly improved the governance model of non-intelligent set-top boxes.
Smart Images

Figure CN118921526A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of digital television, and in particular to a dual-management operation method and system for a non-intelligent set-top box. Background Art
[0002] In recent years, with the continuous development and popularization of digital TV technology, set-top boxes have become an important carrier for home entertainment and information acquisition. Set-top boxes are mainly divided into two categories: smart set-top boxes and non-smart set-top boxes. Compared with smart set-top boxes, non-smart set-top boxes have relatively simple functions and are mainly used to receive and decode digital TV signals. The price is also relatively low, and there is still a large stock market in some rural and remote areas.
[0003] While non-intelligent set-top boxes bring convenience to users, they also have some problems that need to be solved urgently. For example, due to the limitations of hardware configuration and software functions, new applications and services cannot be flexibly expanded, the content sources that users can choose from are relatively limited, and the viewing experience needs to be improved. At the same time, through illegal tampering with firmware and private installation of third-party software, the originally closed system becomes open, making users easily affected and misled by bad information.
[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention
[0005] The embodiment of the present invention provides a dual-management operation method and system for a non-intelligent set-top box, which can at least solve some of the problems existing in the prior art.
[0006] A first aspect of an embodiment of the present invention provides a dual-management operation method for a non-intelligent set-top box, comprising: Based on the end-cloud collaborative architecture, an illegal set-top box identification model is built to obtain communication data packets and extract multi-dimensional features. The key multi-dimensional features are screened with the attention mechanism, and the interactive semantic features are obtained by fusing the key multi-dimensional features. The convolutional neural network is used for joint modeling and deep cross-learning to obtain a comprehensive feature map. An illegal behavior classifier is built and illegal behaviors are identified through small sample learning technology. For illegal set-top boxes, the optimal penalty action sequence is generated based on the reinforcement learning algorithm and automatically triggered through the smart contract program. The user behavior history and social relationship are learned through graph embedding representation through the graph neural network to determine the illegal behavior propagation path and impact chain, identify key propagation nodes, and simulate the effect of cutting off the propagation path based on the causal intervention algorithm to generate the optimal tracing and crackdown plan. The viewing behavior data of non-intelligent set-top boxes is obtained through a big data analysis platform. Based on user portrait technology, the factor decomposition machine and the attention mechanism are combined to extract implicit feature factors and construct user portraits. Based on the user portraits, the viewing preferences and interactive capabilities of each user are predicted in combination with a machine learning algorithm to obtain the upgrade decision parameters corresponding to the current user. Based on the upgrade decision parameters, differentiated upgrade instructions are sent through a remote upgrade server. For the upgraded set-top boxes, digital content and value-added applications are issued in combination with the user portraits. In the big data analysis platform, all data related to the audience rating are integrated and a global knowledge graph is constructed. The entities and relationships in the global knowledge graph are embedded into a low-dimensional semantic space through knowledge representation learning technology, a semantic association network is generated, and the similarities between different user nodes, different content nodes and different advertising nodes are calculated through a collaborative filtering algorithm, a multi-dimensional semantic matching result is generated, and a personalized content recommendation list is determined. Content is delivered according to the personalized content recommendation list, and after the delivery is completed, the effect data per unit time is recorded and the effect data is processed by an online multi-armed bandit algorithm to generate a content governance report, and the content governance report and the optimal traceability and combat plan are integrated to obtain a dual governance report.
[0007] In an optional embodiment, An illegal set-top box recognition model is built based on the end-cloud collaborative architecture. Communication data packets are obtained and multi-dimensional features are extracted. The key multi-dimensional features are screened with the attention mechanism, and the interactive semantic features are obtained by fusing the key multi-dimensional features. The convolutional neural network is used for joint modeling and deep cross-learning to obtain a comprehensive feature map. The illegal behavior classifier is built through small sample learning technology and the illegal behaviors are identified, including: An illegal set-top box identification model is constructed based on a preset end-cloud collaborative architecture, a communication data packet is obtained as input data of the illegal set-top box identification model, link layer attributes, network layer attributes, transport layer attributes, application layer attributes and load content in the communication data packet are extracted, and multi-dimensional heterogeneous data are obtained by combining them. For the multi-dimensional heterogeneous data, frequency features, periodic features and object features corresponding to the communication mode are extracted by statistical analysis methods, and communication mode statistical features are obtained by combining them. For the load content, context features are extracted by natural language processing and sequence mining technology, and combined features corresponding to abnormal behaviors are extracted by combining frequent pattern mining technology. The extracted features are uploaded to the cloud, and pairwise similarity calculations are performed on all feature dimensions to generate a global feature correlation graph. The weights of the nodes and edges in the global feature correlation graph are learned through the attention mechanism to obtain the significance weight of each node and the importance weight of each edge, the global feature correlation graph is pruned and the pruned node features are extracted to obtain the key multidimensional features, the key multidimensional features are mapped to a low-dimensional vector space through a topological association algorithm based on graph embedding, the distances of different key multidimensional features in the low-dimensional vector space are calculated and the low-dimensional space similarity is determined, the temporal features are determined in combination with the behavioral events corresponding to the key multidimensional features, and the temporal causal dependencies between different key multidimensional features are calculated through a causal inference method, the kernel representation of the low-dimensional space similarity and the temporal causal dependencies is learned through a multi-core learning method, and mapped to a common semantic concept space through a multi-view joint dimensionality reduction strategy to generate the interactive semantic features; Based on the interactive semantic features and the key multidimensional features, a convolutional neural network is used for joint modeling to generate an initial comprehensive feature map, the interactive semantic features and the key multidimensional features are used as inputs of the convolutional neural network, the map embedding of the initial comprehensive feature map is used as the convolution kernel, the activation value is determined by a convolution operation, the map embedding is updated by end-to-end back propagation to obtain the comprehensive feature map, based on the node embedding vector corresponding to each node in the comprehensive feature map, a training set is constructed and a small sample classifier is set as an illegal behavior classifier, the hyperparameters of the illegal behavior classifier are randomly initialized and the training set is added to the illegal behavior classifier for training, the hyperparameters of the illegal behavior classifier are repeatedly adjusted until the illegal behavior classifier converges, the trained illegal behavior classifier is used to identify the illegal behavior of the non-intelligent set-top box, and if illegal behavior exists, the current non-intelligent set-top box is marked as an illegal set-top box.
[0008] In an optional embodiment, For illegal set-top boxes, the optimal penalty action sequence is generated based on the reinforcement learning algorithm and automatically triggered through the smart contract program. The user behavior history and social relationships are learned through graph embedding representation through the graph neural network to determine the propagation path and impact chain of illegal behavior, identify key propagation nodes and simulate the effect of cutting off the propagation path based on the causal intervention algorithm, and generate the optimal tracing and crackdown plan, including: For the illegal set-top box, the penalty process is modeled as a Markov decision process through a reinforcement learning algorithm, a state space is set and state parameters are determined, the state parameters include the dynamic structure of the propagation network, the attributes of each node and the global attributes, an action space is set and the penalty intensity and resource consumption corresponding to each action are determined, at each time step, the punished node and the penalty action are selected as a combined action, a state transfer function and a reward function are defined, an action value function is set based on a deep reinforcement learning algorithm and an action with a maximum function value is selected, a value function is fitted through a deep neural network, and the optimal penalty action corresponding to the current time step is solved and obtained, the solution is repeated to obtain the optimal penalty action sequence, the optimal penalty action sequence is written into a smart contract, and the penalty execution is controlled through a contract program; Determine the current violating user based on the optimal penalty action sequence, initialize the user set and the social network of the current violating user according to the comprehensive feature map, obtain the behavior history of the current violating user, perform representation learning on the current violating user based on the graph neural network, aggregate the node information of the current violating user and the node information corresponding to the neighbor nodes of the current violating user in the social network through the graph convolution layer, aggregate the behavior history through the attention mechanism, calculate the corresponding attention weight for each behavior and perform node aggregation to obtain the user embedding representation; Based on the user embedding representation, randomly select a user pair and calculate the user similarity of the current user pair, calculate the corresponding illegal behavior propagation probability according to the user similarity, generate a propagation probability matrix and perform pruning to obtain a directed weighted propagation network, based on the directed weighted propagation network, through a depth-first search algorithm, starting from the node corresponding to the current illegal user, traverse the outgoing edge with the highest propagation probability until reaching the visited node or leaf node, obtain the illegal behavior propagation path, add randomness to the depth-first search algorithm and search again, obtain multiple propagation paths and perform a union operation on the nodes activated on each propagation path to obtain an influence chain, and determine the key propagation node based on the influence chain and the illegal behavior propagation path; A causal graph is constructed based on the causal intervention algorithm, the conditional probability of combating violations is calculated based on the causal graph and the backdoor criterion, the average causal effect is calculated based on the conditional distribution and the key propagation node with the maximum causal effect is selected based on the greedy strategy, the effect of cutting off the propagation path is simulated on the key propagation node and the optimal tracing and combat plan is generated.
[0009] In an optional implementation, the reward function is as shown in the following formula: ; in, r ( s , v , a) indicates the status s Midpoint v In Action a The return value after α 1 represents the change weight, △ N ( s , v , a ) indicates the status s Midpoint v In Action a The change value after α 2 represents the node importance weight, I ( v ) represents a node v The importance of β represents the cost weight, C ( a ) indicates an action a The corresponding cost.
[0010] In an optional embodiment, The viewing behavior data of non-intelligent set-top boxes is obtained through a big data analysis platform. Based on user portrait technology, the factor decomposition machine and the attention mechanism are combined to extract implicit feature factors and construct user portraits. Based on the user portraits, the viewing preferences and interactive capabilities of each user are predicted in combination with a machine learning algorithm to obtain the upgrade decision parameters corresponding to the current user. Based on the upgrade decision parameters, differentiated upgrade instructions are sent through a remote upgrade server. For the upgraded set-top boxes, digital content and value-added applications are sent in combination with the user portraits, including: Based on a pre-set big data analysis platform, the viewing behavior data of non-intelligent set-top boxes are collected and stored, wherein the viewing behavior data includes channel switching, viewing duration and viewing time period, the viewing behavior data, users and programs are represented as triples, the viewing behavior data is modeled by a factor decomposition machine, each user and program is mapped to a low-dimensional latent vector space, the user latent vector and the program latent vector are obtained and the inner product is calculated, the current user's preference for the current program is obtained according to the calculation result, the preference degree is used as a feature interaction term, and the prediction function corresponding to the factor decomposition machine is constructed in combination with the global bias term and the feature weight; For the feature interaction term in the prediction function, the attention mechanism is used to output a scalar score value with the hidden state vector, user hidden vector and program hidden vector at the current moment as data, and the scalar score value corresponding to each time step is indexed and normalized to obtain the feature interaction term weight corresponding to each time step, and the feature interaction term weight is added to the prediction function to solve and obtain the user portrait; Constructing a user feature vector based on the user portrait and the user latent vector, training a preset user classification model through a machine learning algorithm, taking the user feature vector as input, predicting the viewing preference category and interaction ability level of the current user, and obtaining an upgrade decision parameter corresponding to the current user based on the viewing preference category and the interaction ability level; Based on the upgrade decision parameters, a differentiated upgrade instruction is sent to the target set-top box through a remote upgrade server to trigger software upgrade and function update. The differentiated upgrade instruction includes the target software version, upgrade time and upgrade priority. For the upgraded set-top box, combined with the user portrait and the upgrade decision parameters, personalized digital content and value-added applications are matched for the current user through a content filtering algorithm.
[0011] In an optional embodiment, The matching formula through content filtering algorithm is as follows: ; Among them, S represents the similarity, I un Indicates user u and users n A collection of programs that have been interacted with together. f () represents the time decay function, t ui Indicates user u About the show i The timestamp of the interaction. t ni Indicates user n About the show i The timestamp of the interaction. r ui Indicates user u About the show i The preference value of r ni Indicates user n About the show i The preference value of Indicates user u Average preference for the program, Indicates user n Average preference for the program.
[0012] In an optional embodiment, In the big data analysis platform, all the data related to the audience rating are integrated and a global knowledge graph is constructed. The entities and relationships in the global knowledge graph are embedded into a low-dimensional semantic space through knowledge representation learning technology, a semantic association network is generated, and similarities between different user nodes, different content nodes, and different advertising nodes are respectively calculated through a collaborative filtering algorithm, a multi-semantic matching result is generated, and a personalized content recommendation list is determined. Content is delivered according to the personalized content recommendation list, and after delivery, the effect data per unit time is recorded and the effect data is processed by an online multi-armed bandit algorithm to generate a content governance report. The content governance report and the optimal traceability strike plan are integrated to obtain a dual governance report including: Integrate all the data related to the viewing in the big data analysis platform, including users, content, interactions, advertisements and ratings, construct the global knowledge graph based on the integrated data, and for the global knowledge graph, use the relationship between each entity as a translation vector through knowledge representation learning technology, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space to obtain entity semantic vectors, calculate the semantic relevance between different entities through a preset vector space model for the entity semantic vectors, generate a semantic association matrix, and construct a semantic association network based on the semantic association matrix, wherein the nodes in the semantic association network represent entities, and the weighted edges between nodes represent the strength of semantic association between entities; Based on the semantic association network, combined with the collaborative filtering idea, the similarity between different users is calculated based on cosine similarity and a user similarity matrix is generated, the similarity between the target content and other content is calculated and a content similarity matrix is generated, the similarity between the target advertising node and the candidate advertising node is calculated and an advertising similarity matrix is generated, based on the user similarity matrix, the content similarity matrix and the advertising similarity matrix obtained by calculation, the comprehensive matching degree is calculated by weighted fusion to generate the multivariate semantic matching result, based on the multivariate semantic matching result, a mixed integer programming model is constructed in combination with pre-set diversity constraints and novelty constraints, the mixed integer programming model is solved by a branch and bound algorithm to obtain the personalized content recommendation list, and content delivery is performed according to the personalized content recommendation list; Record the effect data per unit time after the delivery is completed and define key performance indicators, calculate the key performance indicators of the personalized content recommendation list based on the effect data, evaluate the recommendation effect based on the key performance indicators, model the content recommendation problem as a multi-armed slot machine problem, use the recommendation strategy corresponding to the personalized content recommendation list as the slot arm, modify the selection probability of each slot arm through an online learning algorithm and improve the recommendation effect, repeatedly optimize to obtain the optimal content governance decision, generate a content governance report based on the optimal content governance decision, and integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
[0013] A second aspect of an embodiment of the present invention provides a non-intelligent set-top box dual-governance operating system, including: The first unit is used to build an illegal set-top box identification model based on the end-cloud collaborative architecture, obtain communication data packets and extract multi-dimensional features, combine the attention mechanism to perform feature screening to obtain key multi-dimensional features, fuse the key multi-dimensional features to obtain interactive semantic features, perform joint modeling and deep cross-learning through convolutional neural networks, obtain a comprehensive feature map, build an illegal behavior classifier and identify illegal behaviors through small sample learning technology, generate the optimal penalty action sequence for illegal set-top boxes based on the reinforcement learning algorithm and automatically trigger it through the smart contract program, and use the graph neural network to perform graph embedding representation learning on user behavior history and social relationships, determine the illegal behavior propagation path and impact chain, identify key propagation nodes, and simulate the effect of cutting off the propagation path based on the causal intervention algorithm to generate the optimal tracing and crackdown plan; The second unit is used to obtain the viewing behavior data of non-intelligent set-top boxes through a big data analysis platform, extract implicit feature factors based on user portrait technology, combine factor decomposition machine and attention mechanism, and construct user portraits, and predict the viewing preference and interaction ability of each user based on the user portraits in combination with machine learning algorithms to obtain the upgrade decision parameters corresponding to the current user, and send differentiated upgrade instructions through a remote upgrade server based on the upgrade decision parameters, and for the upgraded set-top boxes, send digital content and value-added applications in combination with the user portraits; The third unit is used to integrate all the data related to the viewing in the big data analysis platform and construct a global knowledge graph, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space through knowledge representation learning technology, generate a semantic association network, and calculate the similarities between different user nodes, different content nodes and different advertising nodes through a collaborative filtering algorithm, generate multi-dimensional semantic matching results and determine a personalized content recommendation list, deliver content according to the personalized content recommendation list, record the effect data per unit time after the delivery is completed, and process the effect data through an online multi-armed bandit algorithm to generate a content governance report, and integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
[0014] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0015] According to a fourth aspect of the embodiments of the present invention, 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.
[0016] In the present invention, through the end-cloud collaborative architecture and deep learning algorithm, the multi-dimensional characteristics of communication data are fully integrated to build an accurate illegal behavior recognition model, which can effectively detect and identify the tampering and cracking behavior of the set-top box, and provide a basis for timely blocking the spread of illegal content. By using the graph neural network and causal intervention algorithm, the key propagation nodes of illegal behavior can be accurately located, the blocking effects of different crackdown plans can be simulated and evaluated, and the optimal tracing crackdown strategy can be formulated to curb the spread of illegal and harmful information from the source. The optimal punishment action sequence obtained by reinforcement learning is triggered and executed through a smart contract. The immutability and traceability of the blockchain are used to ensure that the punishment process is open, transparent and non-repudiable, thereby improving The credibility of supervision is enhanced by constructing a global knowledge graph covering multi-dimensional data such as users, content, and advertisements, using knowledge representation learning to reveal implicit associations between entities, and combining collaborative filtering algorithms to generate personalized recommendations to meet the differentiated needs of users. In summary, the present invention constructs a full-process, dynamically optimized dual-governance operation method for non-intelligent set-top boxes, which can accurately identify and trace the source of violations, implement upgrade strategies according to local conditions, and optimize content recommendations through multi-dimensional integration to form a joint force for governance and crackdown. It has significant innovation and practicality in purifying the audio-visual environment, standardizing the industry ecology, and ensuring information security, and has revolutionary improvements and enhancements to the traditional non-intelligent set-top box governance model. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 A schematic diagram of a flow chart of a dual management operation method for a non-intelligent set-top box according to an embodiment of the present invention; Figure 2 The figure is a schematic diagram of the structure of the dual-management operating system of a non-intelligent set-top box according to an embodiment of the present invention. DETAILED DESCRIPTION
[0018] 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.
[0019] 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.
[0020] Figure 1FIG. 1 is a flow chart of a dual-management operation method for a non-intelligent set-top box according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Build an illegal set-top box recognition model based on the end-cloud collaborative architecture, obtain communication data packets and extract multi-dimensional features, combine the attention mechanism to perform feature screening to obtain key multi-dimensional features, fuse the key multi-dimensional features to obtain interactive semantic features, perform joint modeling and deep cross-learning through convolutional neural networks, obtain a comprehensive feature map, build an illegal behavior classifier and identify illegal behaviors through small sample learning technology, generate the optimal penalty action sequence for illegal set-top boxes based on the reinforcement learning algorithm and automatically trigger it through the smart contract program, use the graph neural network to perform graph embedding representation learning on user behavior history and social relationships, determine the illegal behavior propagation path and impact chain, identify key propagation nodes, and simulate the effect of cutting off the propagation path based on the causal intervention algorithm to generate the optimal tracing and crackdown plan; The end-cloud collaborative architecture is an architecture that combines edge computing with cloud computing. It processes data at the edge of the network (such as IoT devices) to reduce latency and bandwidth requirements, while using the powerful storage and computing capabilities of cloud computing to perform complex data analysis and processing. The illegal set-top box identification model is a model for detecting and identifying illegal set-top boxes. It determines whether a device is legal based on the device's behavior, network traffic characteristics, and other identification characteristics. The communication data packet is the basic unit in network transmission, which includes the sender, receiver address, and actual data content. The key multidimensional features refer to the most important multi-dimensional features identified in data analysis, which can effectively describe and distinguish data samples. The interactive semantic features are obtained by analyzing user The features extracted from the interactive behaviors between are often used in social networks and recommendation systems to capture the context and meaning of user behaviors. The deep cross-learning is a method of learning interactive features using deep learning models. It is often used in recommendation systems and advertising. It improves prediction accuracy by modeling complex relationships between features. The comprehensive feature map is a map formed by integrating multiple features (such as user behaviors, device features, etc.) to more comprehensively understand and analyze data. The small sample learning is a machine learning method that aims to extract effective information from limited data samples to improve the generalization ability of the model. It is suitable for scenarios with scarce data. The illegal behavior classifier is a machine learning model for identifying and classifying illegal behaviors. It can be based on input The characteristics are used to determine whether the behavior is legal. The optimal penalty action sequence refers to the action sequence that can most effectively reduce improper behavior to achieve the predetermined goal. The smart contract is a contract program that is automatically executed on the blockchain. It can automatically execute the contract terms according to preset conditions to ensure transparent and secure transactions. The graph neural network is a type of deep learning model used to process graph structure data. It can capture the relationship and characteristics between nodes and is used in social networks, recommendation systems and other fields. The user behavior history refers to all the behavior records of users on a certain platform or system, which is used to analyze user preferences, habits and trends. The graph embedding representation learning is a technology that converts graph structure data into low-dimensional vector representations to perform more effective machine learning tasks, such as Classification and clustering. The illegal behavior propagation path refers to the path of illegal behavior spreading on the Internet or social platforms. Analyzing these paths helps to understand and prevent illegal activities. The impact chain describes a series of interrelated events or behaviors, which is usually used to analyze the source of influence and propagation process of a certain behavior or event. The key propagation node refers to the node that plays an important role in the spread of information or behavior. Identifying these nodes helps to carry out effective intervention. The causal intervention algorithm is used to determine the causal impact that may occur when a certain intervention is carried out in a system, so as to evaluate its effectiveness and optimize decisions. The optimal tracing and crackdown plan refers to the most effective plan to track and crack down on the source of illegal activities through analysis and modeling to achieve governance and prevention.
[0021] In an optional embodiment, An illegal set-top box recognition model is built based on the end-cloud collaborative architecture. Communication data packets are obtained and multi-dimensional features are extracted. The key multi-dimensional features are screened with the attention mechanism, and the interactive semantic features are obtained by fusing the key multi-dimensional features. The convolutional neural network is used for joint modeling and deep cross-learning to obtain a comprehensive feature map. The illegal behavior classifier is built through small sample learning technology and the illegal behaviors are identified, including: An illegal set-top box identification model is constructed based on a preset end-cloud collaborative architecture, a communication data packet is obtained as input data of the illegal set-top box identification model, link layer attributes, network layer attributes, transport layer attributes, application layer attributes and load content in the communication data packet are extracted, and multi-dimensional heterogeneous data are obtained by combining them. For the multi-dimensional heterogeneous data, frequency features, periodic features and object features corresponding to the communication mode are extracted by statistical analysis methods, and communication mode statistical features are obtained by combining them. For the load content, context features are extracted by natural language processing and sequence mining technology, and combined features corresponding to abnormal behaviors are extracted by combining frequent pattern mining technology. The extracted features are uploaded to the cloud, and pairwise similarity calculations are performed on all feature dimensions to generate a global feature correlation graph. The weights of the nodes and edges in the global feature correlation graph are learned through the attention mechanism to obtain the significance weight of each node and the importance weight of each edge, the global feature correlation graph is pruned and the pruned node features are extracted to obtain the key multidimensional features, the key multidimensional features are mapped to a low-dimensional vector space through a topological association algorithm based on graph embedding, the distances of different key multidimensional features in the low-dimensional vector space are calculated and the low-dimensional space similarity is determined, the temporal features are determined in combination with the behavioral events corresponding to the key multidimensional features, and the temporal causal dependencies between different key multidimensional features are calculated through a causal inference method, the kernel representation of the low-dimensional space similarity and the temporal causal dependencies is learned through a multi-core learning method, and mapped to a common semantic concept space through a multi-view joint dimensionality reduction strategy to generate the interactive semantic features; Based on the interactive semantic features and the key multidimensional features, a convolutional neural network is used for joint modeling to generate an initial comprehensive feature map, the interactive semantic features and the key multidimensional features are used as inputs of the convolutional neural network, the map embedding of the initial comprehensive feature map is used as the convolution kernel, the activation value is determined by a convolution operation, the map embedding is updated by end-to-end back propagation to obtain the comprehensive feature map, based on the node embedding vector corresponding to each node in the comprehensive feature map, a training set is constructed and a small sample classifier is set as an illegal behavior classifier, the hyperparameters of the illegal behavior classifier are randomly initialized and the training set is added to the illegal behavior classifier for training, the hyperparameters of the illegal behavior classifier are repeatedly adjusted until the illegal behavior classifier converges, the trained illegal behavior classifier is used to identify the illegal behavior of the non-intelligent set-top box, and if illegal behavior exists, the current non-intelligent set-top box is marked as an illegal set-top box.
[0022] The link layer attributes involve the basic characteristics of the network link, including bandwidth, delay, transmission error rate, etc., which mainly affect the transmission performance of data packets in the physical network. The network layer attributes include routing selection, IP address allocation, network topology, etc., which are responsible for transmitting data packets between different networks and ensuring that the data can reach the target smoothly. The transport layer attributes involve the reliability and sequentiality of data transmission, including the use of TCP and UDP protocols, controlling data flow and data integrity. The application layer attributes are characteristics related to specific applications, such as data format, user interaction method and protocol specification, which directly affect user experience and function realization. The load content refers to the content transmitted through the network. The actual content of the transmitted data may include text, audio, video and other information, which is a key element in analyzing communication data. The natural language processing is a technology for processing and analyzing human language. Sequence mining focuses on extracting patterns and regularities from sequence data and is widely used in text analysis and time series prediction. The contextual features refer to information related to a specific situation, which is used to better understand and analyze user behavior or events, and usually include time, location, user attributes, etc. The frequent pattern mining technology is used to identify sets of items or patterns that appear frequently in a data set, and is often used in market basket analysis, user behavior analysis, etc. The pairwise similarity calculation is a technology for measuring the similarity between two objects. , widely used in recommendation systems, cluster analysis, etc. The global feature correlation graph is a graph structure that represents the global relationship between features, which is used to analyze the relevance and influence of features. The significance weight is used to measure the importance of features in a specific task, usually based on its contribution to model performance. The importance weight refers to the relative importance of features in the decision-making process, which helps optimize the feature selection and construction of the model. The topological association algorithm based on graph embedding captures the topological relationship between nodes by embedding graph data into a low-dimensional space. It is used for tasks such as community detection and recommendation systems. The low-dimensional space similarity refers to calculating the similarity between objects in a low-dimensional embedding space. It is often used for data analysis and visualization after dimensionality reduction. The temporal causal dependency focuses on the time order and causal relationship between events, and is used to model dynamic systems and predict future trends. The multi-kernel learning method improves the performance of the model by combining multiple kernel functions, and is suitable for complex classification and regression tasks. The multi-view joint dimensionality reduction strategy uses data from multiple views for dimensionality reduction, retaining information while reducing computational complexity, and is often used for data fusion. The common semantic concept space refers to a conceptual framework shared by multiple semantic entities, which promotes understanding and integration between different data sources. The kernel representation refers to mapping data to a high-dimensional space through a kernel method for more effective classification and regression analysis.
[0023] A data collection module and a feature extraction module are set on the end side to collect communication data packets and extract multi-dimensional heterogeneous features. A feature fusion module, a graph construction module, and an illegal behavior identification module are set on the cloud to fuse the features uploaded on the end side, build a comprehensive feature map, and identify illegal behaviors. The network communication data packet is obtained through the data collection module on the end side, and the link layer, network layer, transport layer, application layer attributes and load content are extracted from the data packet. The frequency, period, and object characteristics of the communication mode are extracted using a statistical analysis method to obtain the statistical characteristics of the communication mode. Natural language processing and sequence mining techniques are used to extract the contextual characteristics of the load content. Frequent pattern mining techniques are used to extract the combined characteristics of abnormal behaviors. For example, the DNS request data packet of a smart set-top box within 24 hours is extracted, and its request frequency, request periodicity, and request domain name are analyzed. Keywords in the HTTPS load are extracted, and recurring abnormal domain name combination patterns are mined. The multi-dimensional heterogeneous features extracted on the client side are uploaded to the cloud, and pairwise similarity calculations are performed on all feature dimensions to construct a global feature correlation graph, in which nodes are features and edges are similarities. The attention mechanism is used to learn the weights of nodes and edges in the global feature correlation graph. The global feature correlation graph is pruned based on the learned weights to extract key features. The key features are mapped to a low-dimensional vector space using a graph embedding algorithm, and spatial similarity is calculated. The temporal causal dependency between features is calculated through causal inference in combination with the behavioral events corresponding to the key features. The kernel representation of spatial similarity and temporal causal dependency is learned using a multi-kernel learning method. The kernel representation is mapped to a common semantic concept space using a multi-view joint dimensionality reduction strategy to obtain interactive semantic features. For example, the extracted features such as communication frequency, domain name, and keywords are uploaded to the cloud, and the Pearson correlation coefficient between different features is calculated to construct a feature correlation graph. Key features such as abnormal domain name combinations and DGA domain names are extracted from the global correlation graph, and their cosine similarity in the embedding space is calculated. The temporal causal relationship between domain name requests and DNS responses is mined, and finally high-level semantic features such as "suspicious DNS behavior" are obtained in the semantic concept space. The interactive semantic features and key multidimensional features are input into the convolutional neural network to generate an initial comprehensive feature map. The map embedding of the initial comprehensive feature map is used as the convolution kernel. The map embedding is optimized through convolution operations and back propagation. The optimization is continuously iterated until the comprehensive feature map converges. A training set is constructed based on the node embedding vectors in the comprehensive feature map. A small sample classifier (such as a support vector machine) is set as an illegal behavior classifier. The illegal behavior classifier is trained using the training set. The hyperparameters are optimized through grid search and other techniques. The classifier is repeatedly trained until convergence. The trained illegal behavior classifier is used to identify illegal behaviors of new smart set-top boxes. If the classifier identifies illegal behaviors, the set-top box is removed. Marked as an illegal set-top box, illustratively, with interactive semantic features such as "suspicious DNS behavior" and key features such as abnormal domain name combinations as input, the graph convolutional neural network is used to learn the high-order interactive relationship between different features, and a comprehensive feature map that integrates semantic information and spatiotemporal behavior information is formed. A batch of known normal nodes and illegal nodes are sampled from the comprehensive feature map, and a binary classification support vector machine model is trained using their embedded vectors. The model hyperparameters are optimized using cross-validation. When a new set-top box is connected to the network, its communication behavior features are extracted and a comprehensive feature map is constructed. The trained SVM classifier is used for prediction. If the prediction is illegal behavior, an illegal set-top box alarm is generated.
[0024] In this embodiment, by extracting multiple protocol layer attributes and contents of communication data packets, a comprehensive feature system covering link, network, transmission, and application layer behaviors is constructed, which can comprehensively characterize the behavior pattern of the set-top box. By applying data analysis technologies such as statistical analysis, natural language processing, and frequent pattern mining, effective features can be automatically extracted from massive data, greatly reducing the cost of manual analysis. Through attention mechanisms, graph embedding, causal inference and other technologies, not only the statistical characteristics of the features themselves are considered, but also the spatial correlation and temporal dependency between features are learned to construct a semantically rich comprehensive feature map. In order to solve the problem of sparse illegal behavior samples, Using a small sample classifier and building a training set based on graph embedding can effectively alleviate the problem of insufficient samples and improve the generalization performance of illegal behavior recognition. In summary, this embodiment can process newly connected set-top box data in real time, quickly identify newly emerging illegal behaviors, meet real-time supervision needs, make full use of multiple dimensions of heterogeneous data, and apply a variety of machine learning and data mining technologies. It can better solve the technical challenges faced in illegal set-top box identification, such as data sparsity, difficult behavior characterization, and sample imbalance. It has strong automated intelligent analysis capabilities, can efficiently, accurately, and comprehensively discover illegal set-top box behaviors, and provide strong support for related supervision.
[0025] In an optional embodiment, For illegal set-top boxes, the optimal penalty action sequence is generated based on the reinforcement learning algorithm and automatically triggered through the smart contract program. The user behavior history and social relationships are learned through graph embedding representation through the graph neural network to determine the propagation path and impact chain of illegal behavior, identify key propagation nodes and simulate the effect of cutting off the propagation path based on the causal intervention algorithm, and generate the optimal tracing and crackdown plan, including: For the illegal set-top box, the penalty process is modeled as a Markov decision process through a reinforcement learning algorithm, a state space is set and state parameters are determined, the state parameters include the dynamic structure of the propagation network, the attributes of each node and the global attributes, an action space is set and the penalty intensity and resource consumption corresponding to each action are determined, at each time step, the punished node and the penalty action are selected as a combined action, a state transfer function and a reward function are defined, an action value function is set based on a deep reinforcement learning algorithm and an action with a maximum function value is selected, a value function is fitted through a deep neural network, and the optimal penalty action corresponding to the current time step is solved and obtained, the solution is repeated to obtain the optimal penalty action sequence, the optimal penalty action sequence is written into a smart contract, and the penalty execution is controlled through a contract program; Determine the current violating user based on the optimal penalty action sequence, initialize the user set and the social network of the current violating user according to the comprehensive feature map, obtain the behavior history of the current violating user, perform representation learning on the current violating user based on the graph neural network, aggregate the node information of the current violating user and the node information corresponding to the neighbor nodes of the current violating user in the social network through the graph convolution layer, aggregate the behavior history through the attention mechanism, calculate the corresponding attention weight for each behavior and perform node aggregation to obtain the user embedding representation; Based on the user embedding representation, randomly select a user pair and calculate the user similarity of the current user pair, calculate the corresponding illegal behavior propagation probability according to the user similarity, generate a propagation probability matrix and perform pruning to obtain a directed weighted propagation network, based on the directed weighted propagation network, through a depth-first search algorithm, starting from the node corresponding to the current illegal user, traverse the outgoing edge with the highest propagation probability until reaching the visited node or leaf node, obtain the illegal behavior propagation path, add randomness to the depth-first search algorithm and search again, obtain multiple propagation paths and perform a union operation on the nodes activated on each propagation path to obtain an influence chain, and determine the key propagation node based on the influence chain and the illegal behavior propagation path; A causal graph is constructed based on the causal intervention algorithm, the conditional probability of combating violations is calculated based on the causal graph and the backdoor criterion, the average causal effect is calculated based on the conditional distribution and the key propagation node with the maximum causal effect is selected based on the greedy strategy, the effect of cutting off the propagation path is simulated on the key propagation node and the optimal tracing and combat plan is generated.
[0026] The penalty intensity refers to the intensity of the penalty imposed on improper behavior in reinforcement learning or control strategies, which is used to adjust the behavior of the agent to encourage it to follow the predetermined strategy. The resource consumption refers to the resources required to perform a certain operation or task (such as time, computing power, memory, etc.). In the optimization algorithm, performance and resource utilization efficiency need to be weighed. The penalty action is an action taken against non-compliant behavior, which aims to correct or reduce the occurrence of the behavior, such as restrictions or penalties on illegal users. The reward function is an important concept in reinforcement learning, which is used to evaluate the benefits of the agent after taking an action in a specific state and guide the agent to learn the optimal strategy. The action value function (Q function) refers to the expected return obtained by performing a specific action in a certain state, which is used to evaluate and select the best action. The value function fitting refers to approximating the value function through algorithms (such as linear regression, neural networks, etc.) to enable more effective strategy optimization and decision-making. The optimal penalty action refers to the action that most effectively reduces improper behavior in a specific situation, which is usually determined based on the learning results of the model. The optimal penalty action sequence refers to the action sequence that can maximize positive returns or minimize penalties in a series of operations, which is used to formulate strategies. The illegal user refers to the user who performs in the system Users who violate regulations or engage in non-compliant behaviors usually need to be managed through monitoring and punishment mechanisms. The user pairs refer to the interactive relationships between users. By analyzing these relationships, we can understand the dynamics of behavior propagation and social networks. The propagation probability matrix is used to describe the probability of information propagation between nodes in the network, helping to predict the diffusion path of information or behavior. The directed weighted propagation network refers to a network structure with directionality and weight, which can represent the strength of the relationship between nodes and the direction of information flow. The depth-first search (DFS) algorithm is an algorithm for traversing or searching a tree or graph by going deep into each branch until the target is found or there are no more nodes. The leaf node is the terminal node in the tree or graph structure, without child nodes, and is usually used to represent the final result or a specific instance of data. The greedy strategy is a decision-making method that selects the current optimal option at each step. Although this strategy may not be the best in the global optimal solution, it is usually simple and efficient. The causal graph is used to represent the causal relationship between variables, and helps analyze and reason about the causal relationship in complex systems through graphical methods. The backdoor criterion is used to determine whether there are confounding variables in causal inference to ensure that the inference of causal relationships is valid, which usually involves identifying potential interference factors.
[0027] The penalty process is modeled as a Markov decision process, and the state space is set. The state parameters include the dynamic structure, node attributes and global attributes of the propagation network. The action space is set. Each action corresponds to a certain penalty intensity and resource consumption. At each time step, the punished node and the penalty action are combined into a combined action. The state transfer function and the reward function are defined. The deep reinforcement learning algorithm is used to set the action value function and select the action with the maximum function value. The value function is fitted using a deep neural network to solve the optimal penalty action for the current time step. The above process is repeated to obtain a complete optimal penalty action sequence. The optimal penalty action sequence is written into the smart contract, and the penalty is automatically executed through the contract program. For example, the state space includes parameters such as the degree distribution and clustering coefficient of the propagation network, and the action space includes penalty measures such as banning accounts and restricting propagation. At each time step, the reinforcement learning agent selects the optimal penalty node and penalty action according to the current state of the propagation network, such as banning the illegal account with the strongest propagation ability. After multiple iterations, an optimal penalty action sequence is obtained, such as first restricting propagation and then gradually banning key nodes, which is written into the blockchain contract for automatic execution. Determine the current illegal user from the optimal penalty action sequence, initialize the user set and the social network of the current illegal user according to the comprehensive feature map, obtain the historical behavior data of the current illegal user, use the graph neural network to learn the representation of the current illegal user, aggregate the node information of the current illegal user and the information of the neighbor nodes in the social network through the graph convolution layer, aggregate the behavior history of the current illegal user through the attention mechanism, calculate the attention weight of each behavior, aggregate the results of the graph convolution and attention aggregation, and obtain the user embedding representation vector. For example, determine the illegal user A who needs to be focused on from the penalty action sequence, initialize A's social network according to the user portrait database, including his friends, group chats, etc., perform graph convolution on A's personal attributes and friend attributes through the graph convolution neural network model, perform attention-weighted summation on the dynamics, comments and other behaviors posted by A, and finally obtain an embedding vector that comprehensively represents A's behavior characteristics; Based on the user embedding representation, a batch of user pairs are randomly selected. For each user pair, their user similarity is calculated, such as the cosine similarity of the embedding vector. According to the user similarity matrix, the propagation probability matrix of illegal behaviors between users is calculated. The propagation probability matrix is pruned to obtain a directed weighted graph as a propagation network. For example, 10,000 pairs of users are randomly sampled based on the embedding vector, the similarity of these user pairs is calculated, and the propagation probability matrix is constructed. It is assumed that there is a large propagation probability between users with a similarity greater than 0.8. The propagation network is obtained by pruning, and the weight of the edge is the propagation probability. Starting from the current illegal user node, prioritize traversing the outgoing edges with the highest propagation probability, repeat the above process until reaching the visited node or leaf node, and obtain a complete propagation path. Introduce randomness in the depth-first search process, run multiple times to obtain multiple propagation paths, perform union operations on multiple propagation paths, and obtain a complete influence chain. According to the influence chain and propagation path, determine the key propagation nodes, such as nodes that appear on multiple paths, use the causal intervention algorithm to build a causal graph from punishment actions to illegal propagation. Based on the causal graph and the backdoor criterion formula, calculate the conditional probability of illegal propagation for different crackdown actions. Calculate the average causal effect of each key propagation node after being hit through the causal effect formula, and use The greedy strategy selects the key nodes with the largest causal effect for key strikes, simulates the effect of cutting off the propagation path at the key nodes, and generates the optimal tracing strike plan. For example, on the propagation network, starting from the illegal user A, 10 depth-first searches are run to obtain 10 propagation paths. The nodes activated by these paths are taken as the union to obtain the influence chain. It is found that users B and C frequently appear on the key paths and may be key propagation nodes. A causal graph is constructed based on historical strike and propagation data, and the conditional propagation probability of banning users B and C is calculated. It is found that the average causal effect of banning user B is greater. The cutting effect is simulated at the B node first, and it is found that most of the propagation paths can be blocked, and it is included in the optimal strike plan.
[0028] In this embodiment, by modeling the punishment process as a Markov decision process, the deep reinforcement learning algorithm is used to automatically learn the optimal punishment action sequence. While considering the effects and costs of different punishment measures, the strategy can be adjusted in real time according to the dynamic changes of the propagation network. It is more flexible and intelligent than the traditional static punishment scheme. The social network of the illegal user is represented by a graph neural network, and multi-dimensional heterogeneous information such as user attributes, behavior history, and network structure are fused and encoded into the user embedding vector to obtain a more accurate and comprehensive user portrait, which lays the foundation for subsequent analysis. The propagation network is constructed based on the user embedding vector, and the improved depth-first search algorithm is used to automatically mine multiple propagation paths and extract the influence of common activation. The chain can quickly lock potential key propagation nodes, greatly reducing the cost of manual tracing. By constructing a causal graph and calculating conditional probabilities, the causal effects of different crackdown actions are quantitatively evaluated, and the greedy strategy is used to automatically select the optimal intervention node, making the crackdown plan more causally explanatory and intervention effective. The effect of cutting off the propagation path is simulated at the key nodes, and the optimal tracing crackdown plan is automatically generated. The propagation channels of illegal information are blocked through targeted blocking, flow limiting and other measures, and the spread of illegal activities is curbed from the source. In summary, this embodiment can well cope with dynamic games in complex social network environments, improve the accuracy and timeliness of crackdowns, and provide new technical means and ideas for curbing the spread of illegal information and purifying the network ecology.
[0029] In an optional implementation, the reward function is as shown in the following formula: ; in, r ( s , v , a ) indicates the status s Midpoint v In Action a The return value after α 1 represents the change weight, △ N ( s , v , a ) indicates the status s Midpoint v In Action a The change value after α 2 represents the node importance weight, I ( v ) represents a node v The importance of β represents the cost weight, C ( a ) indicates an action a The corresponding cost.
[0030] In this embodiment, the reward function guides the intelligent agent to select a strategy combination that can maximize the improvement of node status, enhance node importance and minimize action costs. Through weighted balance, the intelligent agent can optimize the status of a single node while taking into account the health of the entire network, avoiding excessive focus on local interests and damaging global performance. By setting cost weights, the frequency and intensity of the action taken by the intelligent agent can be controlled, the negative impact of excessive intervention on the system can be avoided, and the cost-effectiveness of intervention can be improved. By adjusting the change weights, node importance weights and cost weights, the intelligent agent can learn targeted strategies under different task requirements, such as focusing on node status improvement, key node protection or cost control. In summary, this embodiment comprehensively considers node status changes, importance and action costs, and provides a flexible, multi-objective optimization mechanism that can guide the intelligent agent to learn the optimal strategy that balances local benefits and global impacts, which can significantly improve the efficiency and strategy quality of reinforcement learning and accelerate the process of network anomaly detection and intervention optimization.
[0031] S2. Obtain the viewing behavior data of non-intelligent set-top boxes through the big data analysis platform, extract implicit feature factors based on user portrait technology, combine factor decomposition machine and attention mechanism, and construct user portraits. Based on the user portraits, combine machine learning algorithms to predict the viewing preferences and interactive capabilities of each user to obtain the upgrade decision parameters corresponding to the current user. Based on the upgrade decision parameters, send differentiated upgrade instructions through the remote upgrade server. For the upgraded set-top boxes, send digital content and value-added applications based on the user portraits. The user portrait technology is to construct a comprehensive description of the user by collecting and analyzing the user's behavioral data, preferences and characteristics, so as to provide personalized services and recommendations. The factor decomposition machine is a recommendation algorithm that can effectively process high-dimensional sparse data and capture the interaction effects between features by decomposing feature vectors. It is widely used in recommendation systems. The implicit feature factor refers to the unobservable features obtained through learning in the factor decomposition machine, which helps to explain user behavior and preferences. The viewing preference refers to the user's preference for a specific program or content, which can be quantified through historical viewing data and behavioral analysis. The interactive ability refers to the frequency and quality of interaction between users and platforms or content, usually including comments, sharing, likes, etc. The behavior reflects the user's activity. The upgrade decision parameter is a key indicator used to evaluate and decide on content or function upgrades to ensure that the upgrade can meet user needs and improve the experience. The remote upgrade server refers to a server used to push software or content updates to user devices, usually supporting online upgrades and automatic update functions. The differentiated upgrade instructions are personalized upgrade plans formulated according to different user characteristics and needs to ensure that each user obtains the most suitable upgrade content. The digital content refers to various media resources in digital form, such as video, audio, text, etc., which can be distributed and consumed on the network platform. The value-added application refers to providing additional functions or services on top of the basic services to enhance user experience and increase user stickiness.
[0032] In an optional embodiment, The viewing behavior data of non-intelligent set-top boxes is obtained through a big data analysis platform. Based on user portrait technology, the factor decomposition machine and the attention mechanism are combined to extract implicit feature factors and construct user portraits. Based on the user portraits, the viewing preferences and interactive capabilities of each user are predicted in combination with a machine learning algorithm to obtain the upgrade decision parameters corresponding to the current user. Based on the upgrade decision parameters, differentiated upgrade instructions are sent through a remote upgrade server. For the upgraded set-top boxes, digital content and value-added applications are sent in combination with the user portraits, including: Based on a pre-set big data analysis platform, the viewing behavior data of non-intelligent set-top boxes are collected and stored, wherein the viewing behavior data includes channel switching, viewing duration and viewing time period, the viewing behavior data, users and programs are represented as triples, the viewing behavior data is modeled by a factor decomposition machine, each user and program is mapped to a low-dimensional latent vector space, the user latent vector and the program latent vector are obtained and the inner product is calculated, the current user's preference for the current program is obtained according to the calculation result, the preference degree is used as a feature interaction term, and the prediction function corresponding to the factor decomposition machine is constructed in combination with the global bias term and the feature weight; For the feature interaction term in the prediction function, the attention mechanism is used to output a scalar score value with the hidden state vector, user hidden vector and program hidden vector at the current moment as data, and the scalar score value corresponding to each time step is indexed and normalized to obtain the feature interaction term weight corresponding to each time step, and the feature interaction term weight is added to the prediction function to solve and obtain the user portrait; Constructing a user feature vector based on the user portrait and the user latent vector, training a preset user classification model through a machine learning algorithm, taking the user feature vector as input, predicting the viewing preference category and interaction ability level of the current user, and obtaining an upgrade decision parameter corresponding to the current user based on the viewing preference category and the interaction ability level; Based on the upgrade decision parameters, a differentiated upgrade instruction is sent to the target set-top box through a remote upgrade server to trigger software upgrade and function update. The differentiated upgrade instruction includes the target software version, upgrade time and upgrade priority. For the upgraded set-top box, combined with the user portrait and the upgrade decision parameters, personalized digital content and value-added applications are matched for the current user through a content filtering algorithm.
[0033] The low-dimensional latent vector space refers to the process of mapping high-dimensional data to low-dimensional space in order to simplify calculation and analysis while retaining important information. The user latent vector is a low-dimensional vector representing user characteristics and preferences, which is usually generated by a machine learning algorithm to facilitate similarity calculation and recommendation. The program latent vector is a low-dimensional vector representing the characteristics and attractiveness of a program, which is used to capture the relationship between program content and user preferences. The feature interaction term refers to the combined effect between multiple features in the recommendation model, and the prediction accuracy is improved by modeling interaction terms. The global bias term is a constant used to adjust the model prediction results, reflecting the average trend of the overall data, and is often used to improve The accuracy of the model. The prediction function is a function used to generate prediction results based on input features. In the recommendation system, it is used to predict the user's preference for specific content. The latent state vector is a vector that represents the potential state in the model, which is usually used to capture the temporal changes and potential characteristics of user behavior. The scalar score value is a single value output by the model, which is used to represent the user's preference for a certain content or the prediction result. The viewing preference category is to classify the user's viewing preferences in order to better understand and predict user behavior. The interactive ability level is a quantitative assessment of the user's participation and interactive behavior on the platform, which is usually used to divide the user's activity level and influence.
[0034] Based on the pre-built big data analysis platform, the viewing behavior data of non-intelligent set-top boxes are collected and stored. The viewing behavior data mainly includes information such as channel switching, viewing duration and viewing time period. The viewing behavior data, users and programs are represented as triples, such as (user ID, program ID, viewing duration). The viewing behavior data is modeled using a factorization machine. Each user and program is mapped to a low-dimensional latent vector space through the factorization machine to obtain the user latent vector and the program latent vector. The inner product of the user latent vector and the program latent vector is calculated to obtain the current user's preference for the current program. The degree of preference is taken as the feature interaction term, and the prediction function of the factor decomposition machine is constructed by combining the global bias term and the feature weight. For example, the viewing logs of 1 million non-intelligent set-top boxes in one month are collected, and 100 million viewing records of the form (U001, P123, 1800) are obtained. The factor decomposition machine is used for modeling. Assuming that the dimension of the latent vector is 10, the model is trained by the gradient descent method to obtain the 10-dimensional latent vector of each user and program. The inner product of the latent vector of user U001 and program P123 is calculated to obtain the preference degree of 0.8, which is substituted into the prediction function for training; For the feature interaction term in the prediction function, an attention mechanism is introduced for optimization. The hidden state vector, user hidden vector and program hidden vector at the current moment are used as input data. The attention score is calculated through the attention network (such as scaled dot product attention), and a scalar weight is output. The attention weight obtained at each time step is indexed and normalized. The normalized attention weight is used as the dynamic weight of the feature interaction term and added to the prediction function. The prediction function is solved to obtain a user portrait representation that takes into account temporal attention. For example, based on the user program hidden vector obtained in the previous step, for each viewing record of user U001, the hidden state vector, user hidden vector and program hidden vector at the viewing moment are aggregated into a long vector and input into the attention network to obtain an attention weight of 0.6 between 0 and 1. The attention weights of all records of user U001 are normalized, and the final feature interaction term is obtained by weighted average. It is substituted into the prediction function to obtain a 10-dimensional user portrait vector. Based on the user portrait and user latent vector obtained in the previous step, a comprehensive user feature vector is constructed. A user classification model is trained using a machine learning algorithm (such as logistic regression, decision tree, etc.). The user feature vector is used as the model input to predict the current user's viewing preference category (such as sports, movies, etc.), and the current user's interactive ability level (such as high interaction, medium interaction, low interaction). The predicted viewing preference category and interactive ability level are used as parameters for the upgrade decision. Based on the upgrade decision parameters obtained in the previous step, differentiated upgrade instructions are sent to the target set-top box through the remote upgrade server. The upgrade instructions include information such as the target software version, upgrade time, and upgrade priority. For different user groups, the upgrade instructions may be different. For example, high-interaction users are given priority to upgrade enhanced interactive functions. After receiving the upgrade instruction, the target set-top box triggers software upgrades and function updates. Process, for the upgraded set-top box, combined with the user portrait and upgrade decision parameters, the content filtering algorithm is used for personalized recommendation, and the most relevant digital content and value-added applications are matched to the user according to the user's viewing preference to improve the user experience. For example, the 10-dimensional portrait vector and 10-dimensional latent vector of user U001 are combined into a 20-dimensional feature vector, which is input into a three-layer fully connected network for binary and ternary classification. The prediction results are "like sports programs" and "high interactive users", which are used as parameters for the upgrade decision of user U001. According to the "high interactive" attribute of user U001, it is included in the priority upgrade batch, and the latest interactive interface and voice search function are pushed to its set-top box. After the upgrade is completed, based on U001's preference of "like sports", the collaborative filtering algorithm is used to recommend live sports events, sports information and other content to it, and push value-added applications such as sports guessing games.
[0035] In this embodiment, the viewing behavior data is modeled through a factorization machine, and the user and the program are mapped to the same latent vector space, the user's implicit preferences are mined, and a comprehensive, dynamic, and personalized user portrait is generated in combination with user attributes and interactive behaviors to provide a basis for subsequent analysis and decision-making. Machine learning classification can mine more complex user feature patterns, improve the accuracy of classification, provide decision support for differentiated upgrades, and push different upgrade instructions to different groups, such as giving priority to upgrading interactive functions and customizing software versions, etc., to improve the pertinence and effectiveness of upgrades and avoid the inconvenience caused to users by "one-size-fits-all" upgrades. Personalized recommendations can significantly improve user satisfaction and stickiness, extend viewing time, explore potential value-added points, and achieve refined operations. In summary, this embodiment can effectively improve user experience and extend the life cycle of equipment through data-driven user insights and personalized services, and can also help operators achieve refined operations and differentiated competition, and create new value growth points in the stock market. It has broad application prospects and commercial potential.
[0036] In an optional embodiment, The matching formula through content filtering algorithm is as follows: ; Among them, S represents the similarity, I un Indicates user u and users n A collection of programs that have been interacted with together. f () represents the time decay function, t ui Indicates user u About the show i The timestamp of the interaction. t ni Indicates user n About the show i The timestamp of the interaction. r ui Indicates user u About the show i The preference value of r ni Indicates user n About the show i The preference value of Indicates user u Average preference for the program, Indicates user n Average preference for the program.
[0037] In this embodiment, by introducing the time decay function, the algorithm can assign different weights to interactive behaviors at different times, with recent behaviors having higher weights, thereby capturing the dynamic changes of user preferences and improving the timeliness of recommendations. By introducing the time decay function in the similarity calculation formula, the weights of interactive behaviors at different times can be dynamically adjusted, making the similarity calculation more accurate and better reflecting the true similarity of users. By considering the time dimension, the coverage of user interactive behaviors is expanded, which alleviates the data sparsity problem to a certain extent and improves the reliability of similarity calculation. After introducing time decay, the niche preferences of users in different time periods can be mined, the recommendation accuracy of long-tail programs can be improved, and the personalized needs of users can be met. In summary, this embodiment can mine the unique preferences of users in different time periods, provide more diversified and personalized program recommendations, and at the same time alleviate the data sparsity problem and improve the recommendation quality of long-tail programs. The time decay mechanism enhances the interpretability and transparency of recommendations, helps to improve users' trust and satisfaction with the recommendation system, and can significantly improve the user experience and operation effects of non-intelligent set-top boxes.
[0038] S3. Integrate all the data related to the audience rating in the big data analysis platform and construct a global knowledge graph. Through the knowledge representation learning technology, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space, generate a semantic association network, and calculate the similarities between different user nodes, different content nodes and different advertising nodes through a collaborative filtering algorithm. Generate multi-dimensional semantic matching results and determine a personalized content recommendation list. Deliver content according to the personalized content recommendation list. After delivery, record the effect data per unit time and process the effect data with an online multi-armed bandit algorithm to generate a content governance report. Integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
[0039] The global knowledge graph is a graphical structure containing extensive information and relationships, used to represent entities in the real world and the connections between them, and supports reasoning, query and knowledge discovery. The low-dimensional semantic space refers to mapping high-dimensional semantic information to a low-dimensional space through dimensionality reduction technology to facilitate more efficient calculation and analysis while maintaining important semantic relationships. The semantic association network is a network structure representing different entities and their semantic relationships, used to capture and analyze the connections and influences between knowledge. The collaborative filtering algorithm is a recommendation system technology that predicts users' preferences for new content by analyzing their historical behaviors and similar behaviors of other users. The multi-dimensional semantic matching structure The result refers to the result of semantic matching in multiple dimensions, which is usually used to improve the accuracy of information retrieval and recommendation systems. The personalized content recommendation list is a content list generated based on the user's historical behavior, preferences and characteristics, aiming to improve user satisfaction and engagement. The content delivery is the process of pushing specific content to target users, usually based on user characteristics and behavior data to achieve the best effect. The effect data is an indicator used to evaluate the effect of content delivery, such as click-through rate, conversion rate, etc., to help optimize subsequent delivery strategies. The online multi-armed bandit algorithm is a problem for dynamic decision-making, mainly used to select the best option in an uncertain environment to maximize benefits.
[0040] In an optional embodiment, In the big data analysis platform, all the data related to the audience rating are integrated and a global knowledge graph is constructed. The entities and relationships in the global knowledge graph are embedded into a low-dimensional semantic space through knowledge representation learning technology, a semantic association network is generated, and similarities between different user nodes, different content nodes, and different advertising nodes are respectively calculated through a collaborative filtering algorithm, a multi-semantic matching result is generated, and a personalized content recommendation list is determined. Content is delivered according to the personalized content recommendation list, and after delivery, the effect data per unit time is recorded and the effect data is processed by an online multi-armed bandit algorithm to generate a content governance report. The content governance report and the optimal traceability strike plan are integrated to obtain a dual governance report including: Integrate all the data related to the viewing in the big data analysis platform, including users, content, interactions, advertisements and ratings, construct the global knowledge graph based on the integrated data, and for the global knowledge graph, use the relationship between each entity as a translation vector through knowledge representation learning technology, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space to obtain entity semantic vectors, calculate the semantic relevance between different entities through a preset vector space model for the entity semantic vectors, generate a semantic association matrix, and construct a semantic association network based on the semantic association matrix, wherein the nodes in the semantic association network represent entities, and the weighted edges between nodes represent the strength of semantic association between entities; Based on the semantic association network, combined with the collaborative filtering idea, the similarity between different users is calculated based on cosine similarity and a user similarity matrix is generated, the similarity between the target content and other content is calculated and a content similarity matrix is generated, the similarity between the target advertising node and the candidate advertising node is calculated and an advertising similarity matrix is generated, based on the user similarity matrix, the content similarity matrix and the advertising similarity matrix obtained by calculation, the comprehensive matching degree is calculated by weighted fusion to generate the multivariate semantic matching result, based on the multivariate semantic matching result, a mixed integer programming model is constructed in combination with pre-set diversity constraints and novelty constraints, the mixed integer programming model is solved by a branch and bound algorithm to obtain the personalized content recommendation list, and content delivery is performed according to the personalized content recommendation list; Record the effect data per unit time after the delivery is completed and define key performance indicators, calculate the key performance indicators of the personalized content recommendation list based on the effect data, evaluate the recommendation effect based on the key performance indicators, model the content recommendation problem as a multi-armed slot machine problem, use the recommendation strategy corresponding to the personalized content recommendation list as the slot arm, modify the selection probability of each slot arm through an online learning algorithm and improve the recommendation effect, repeatedly optimize to obtain the optimal content governance decision, generate a content governance report based on the optimal content governance decision, and integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
[0041] The knowledge representation learning technology is a technology used to convert information in the knowledge graph into a vector representation to facilitate machine learning and reasoning tasks. The translation vector is an offset used to represent a specific concept or entity, and is often used in knowledge representation learning to capture the relationship between entities. The vector space model is a mathematical model used to represent text or other data. Similarity and relevance are calculated by converting data into vector representations. The semantic relevance refers to the degree of semantic connection between two concepts or entities, which is usually measured by analyzing context or co-occurrence relationships. The semantic association matrix is a matrix used to represent the semantic relationship between multiple entities, which facilitates the analysis and calculation of the relevance between them. The user similarity matrix is a matrix used to represent the similarity between users, which is usually calculated through user behavior and preferences and used in recommendation systems. The content similarity matrix is used A matrix representing the similarity between different contents facilitates content recommendation and retrieval. The advertising similarity matrix is a matrix used to represent the similarity between different advertisements, which helps optimize advertising delivery strategies. The diversity constraint is a constraint in the recommendation system that ensures that there are certain differences between recommended contents to improve user experience and satisfaction. The novelty constraint aims to recommend new and unexplored content to enhance user exploration and satisfaction. The mixed integer programming model is an optimization model that combines integers and continuous variables and is widely used in resource allocation and scheduling problems. The branch and bound algorithm is an algorithm for solving combinatorial optimization problems. It explores and prunes branches through branch trees to find the optimal solution. The slot machine arm is a concept in the online multi-armed slot machine algorithm, representing different options or strategies available, and each "arm" has a corresponding potential benefit.
[0042] In the big data analysis platform, all data related to viewing are collected and integrated, including user attribute data, program content metadata, user-content interaction data, advertising data and ratings statistics, etc. Based on the integrated heterogeneous data, a global knowledge graph is constructed. The nodes in the graph represent various entities (such as users, programs, advertisements, etc.), and the edges represent the associations between entities (such as the user-program viewing relationship, the program-advertising placement relationship, etc.). For example, assume that the following data is collected: user A is a 30-year-old male and a teacher. User A watched news program B, documentary C and movie D. Car advertisement E was placed in program B, and the ratings of program B were 3%. A simple knowledge graph is constructed, including nodes such as user A, program B, C, D, advertisement E, and edges such as the viewing relationship between user A and program B and the placement relationship between program B and advertisement E. For the constructed global knowledge graph, the knowledge representation learning technology is used to convert the graph into The entities and relations in the knowledge graph are embedded into a low-dimensional semantic space, each entity is represented as a semantic vector, each relationship is represented as a translation vector, and the semantic information of entities and relationships is characterized by vector operations. Based on the learned entity semantic vectors, a preset vector space model (such as cosine similarity) is used to calculate the semantic relevance between different entities, and a semantic association matrix is generated. A semantic association network is constructed according to the semantic association matrix, in which nodes represent entities, and weighted edges between nodes represent the strength of semantic association between entities. For example, embedding representation learning is performed on entities and relations in the knowledge graph to obtain semantic vectors corresponding to user A, program B, C, D, and advertisement E, and the cosine similarity between them is calculated. It is found that the similarity between user A and programs B and C is high, and the similarity between program B and advertisement E is high. A semantic association network is constructed, in which user A, program B, C, D, and advertisement E are nodes, and the weighted edges between nodes reflect the strength of semantic relevance. Based on the constructed semantic association network, the idea of collaborative filtering is used to calculate similarities from the three dimensions of user, content, and advertisement. The similarity between any two user nodes is calculated to generate a user similarity matrix. The similarity between any two content nodes is calculated to generate a content similarity matrix. The similarity between any two advertisement nodes is calculated to generate an advertisement similarity matrix. The similarity calculation can be based on common methods such as cosine similarity. For example, based on the above semantic association network, the user similarity matrix, the content similarity matrix, and the advertisement similarity matrix are calculated. In the program dimension, it can be calculated that the similarity between program B and program C is 0.8, the similarity between program B and program D is 0.5, and the similarity between program C and program D is 0.6, thereby obtaining a 3x3 content similarity matrix. For a given target user, the similarity and association strength are comprehensively considered from the three dimensions of user, content, and advertisement, and the matching degree between other entities and the target user is calculated. The matching degree of the user dimension is based on the user similarity matrix, the matching degree of the content dimension is based on the content similarity matrix, and the matching degree of the advertisement dimension is based on the advertisement similarity matrix. The matching degrees of the three dimensions are weighted and fused to obtain a comprehensive matching degree as the basis for generating recommendation results. When generating recommendation results, it is also necessary to consider diversity constraints (such as diversity of content types to avoid recommending homogeneous content) and novelty constraints (such as giving priority to recommending new content that users have not interacted with). The recommendation problem is formalized as a mixed integer programming model. The goal is to meet the constraints of diversity and novelty. Under the premise of the bundle, the comprehensive matching degree between the recommended content and the user is maximized, and the mixed integer programming is solved by combinatorial optimization methods such as the branch and bound algorithm to obtain an optimal personalized recommendation list. For example, assuming that personalized recommendations are to be generated for user A, according to the user similarity matrix, it is found that user A and user F are similar, and user F has watched program G and advertisement H. According to the content similarity matrix, it is found that program C and program E have a high similarity, and according to the advertisement similarity matrix, it is found that advertisement E and advertisement I have a high similarity. Comprehensively considering the matching degree of the three dimensions of user, content, and advertisement, and considering the constraints of diversity and novelty, by solving the mixed integer programming, a recommendation list for user A is finally obtained, including programs C, E, G and advertisement I; The personalized recommendation list and its corresponding recommendation strategy are applied to the actual system for content delivery. After a period of delivery, the feedback data of the recommendation is collected, such as user clicks, viewing, ratings and other behaviors, and some key performance indicators are calculated based on this, such as click-through rate, viewing time, user satisfaction, etc., which are used to evaluate the recommendation effect. The entire recommendation problem is modeled as a multi-armed slot machine problem, in which each slot machine arm corresponds to a recommendation strategy. The online learning algorithm based on the multi-armed slot machine (such as Thompson sampling) dynamically adjusts the selection probability of different recommendation strategies, so that the overall recommendation effect is continuously improved. Through continuous online learning and strategy optimization, a recommendation strategy with optimal performance is finally obtained as the decision-making basis for content governance. For example, The obtained recommendation list is delivered to user A, and feedback data within 7 days is collected. It is found that user A clicked on program C and advertisement I, and gave a high rating to program E, but had no interactive behavior on program G. The click-through rate, average viewing time and user satisfaction of the recommendation list are calculated. The strategy corresponding to the recommendation list is regarded as a slot machine arm, and is incorporated into the online learning framework of the multi-armed slot machine together with other candidate recommendation strategies. Through the Thompson sampling algorithm, different strategy combinations are continuously tried, and the selection probability of each strategy is adjusted according to the feedback data. After a period of online learning, it is found that a new recommendation strategy (such as replacing program G in the initial list) can achieve better overall results, so it is used as the optimal decision for content governance; Based on the optimal recommendation strategy obtained from online learning, a complete content governance decision is formed, including which content should be recommended, with what priority and frequency, etc. The optimal decision obtained from learning is summarized into a content governance report as a guide for optimizing content production, delivery, sorting and other links. The content governance report and the optimal traceability and crackdown plan in the previous steps are integrated to form a comprehensive dual governance report, covering the two major governance measures of content recommendation and public opinion supervision, providing a systematic solution for the platform's content ecological governance. For example, based on the optimal recommendation strategy obtained, a content governance report is generated, suggesting that when producing content, the platform should produce more content that is highly matched with user A's preferences, such as programs C and E. When delivering content, programs C, E and advertisement I should be pushed first, and the push frequency of program G should be appropriately reduced. When sorting content, programs C, E, etc. are ranked in the front. The content governance report is summarized with the governance plan for traceability and crackdown, and a comprehensive dual governance report is submitted as an action guide for platform content governance.
[0043] In this embodiment, the implicit semantic associations between users, content and advertisements are mined through the global knowledge graph and representation learning technology, revealing their similarities and consistency in the deep semantic space, overcoming problems such as data sparsity and cold start, and improving the accuracy of personalized recommendations. When generating the recommendation list, diversity constraints are introduced to avoid excessive homogeneity of the recommendation results, ensuring the diversity of the recommended content in terms of type and style, thereby improving user satisfaction and interaction rate. By imposing novelty constraints, new content that users have not interacted with is recommended first, expanding the user's content contact surface, discovering potential points of interest, stimulating the user's motivation to explore new things, and enhancing user stickiness. Based on the online learning algorithm, Dynamically adjust the selection probability of the recommendation strategy, utilize existing effective strategies, and continuously explore new strategy space, continuously optimize in the balance between utilization and exploration, so that the long-term effect of the recommendation system is improved. Through the online learning mechanism, the recommendation system can continuously capture changes in user preferences and adjust the recommendation strategy in time to ensure that the recommendation results are always in line with the user's immediate preferences. In summary, this embodiment achieves the optimization of long-term effects and adaptability, forms a synergistic effect of traffic governance and content governance, and provides a flexible and scalable solution for the platform's full-process content governance. It has significant technical effects and application value in improving user experience, optimizing content ecology, and ensuring information security.
[0044] Figure 2 FIG. 1 is a schematic diagram of the structure of a dual-management operating system for a non-intelligent set-top box according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to build an illegal set-top box identification model based on the end-cloud collaborative architecture, obtain communication data packets and extract multi-dimensional features, combine the attention mechanism to perform feature screening to obtain key multi-dimensional features, fuse the key multi-dimensional features to obtain interactive semantic features, perform joint modeling and deep cross-learning through convolutional neural networks, obtain a comprehensive feature map, build an illegal behavior classifier and identify illegal behaviors through small sample learning technology, generate the optimal penalty action sequence for illegal set-top boxes based on the reinforcement learning algorithm and automatically trigger it through the smart contract program, and use the graph neural network to perform graph embedding representation learning on user behavior history and social relationships, determine the illegal behavior propagation path and impact chain, identify key propagation nodes, and simulate the effect of cutting off the propagation path based on the causal intervention algorithm to generate the optimal tracing and crackdown plan; The second unit is used to obtain the viewing behavior data of non-intelligent set-top boxes through a big data analysis platform, extract implicit feature factors based on user portrait technology, combine factor decomposition machine and attention mechanism, and construct user portraits, and predict the viewing preference and interaction ability of each user based on the user portraits in combination with machine learning algorithms to obtain the upgrade decision parameters corresponding to the current user, and send differentiated upgrade instructions through a remote upgrade server based on the upgrade decision parameters, and for the upgraded set-top boxes, send digital content and value-added applications in combination with the user portraits; The third unit is used to integrate all the data related to the viewing in the big data analysis platform and construct a global knowledge graph, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space through knowledge representation learning technology, generate a semantic association network, and calculate the similarities between different user nodes, different content nodes and different advertising nodes through a collaborative filtering algorithm, generate multi-dimensional semantic matching results and determine a personalized content recommendation list, deliver content according to the personalized content recommendation list, record the effect data per unit time after the delivery is completed, and process the effect data through an online multi-armed bandit algorithm to generate a content governance report, and integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
[0045] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0046] According to a fourth aspect of the embodiments of the present invention, 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.
[0047] 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.
[0048] 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 dual management operation method for a non-intelligent set-top box, characterized in that: include: Based on the end-cloud collaborative architecture, an illegal set-top box identification model is built to obtain communication data packets and extract multi-dimensional features. The key multi-dimensional features are screened with the attention mechanism, and the interactive semantic features are obtained by fusing the key multi-dimensional features. The convolutional neural network is used for joint modeling and deep cross-learning to obtain a comprehensive feature map. An illegal behavior classifier is built and illegal behaviors are identified through small sample learning technology. For illegal set-top boxes, the optimal penalty action sequence is generated based on the reinforcement learning algorithm and automatically triggered through the smart contract program. The user behavior history and social relationship are learned through graph embedding representation through the graph neural network to determine the illegal behavior propagation path and impact chain, identify key propagation nodes, and simulate the effect of cutting off the propagation path based on the causal intervention algorithm to generate the optimal tracing and crackdown plan. The viewing behavior data of non-intelligent set-top boxes is obtained through a big data analysis platform. Based on user portrait technology, the factor decomposition machine and the attention mechanism are combined to extract implicit feature factors and construct user portraits. Based on the user portraits, the viewing preferences and interactive capabilities of each user are predicted in combination with a machine learning algorithm to obtain the upgrade decision parameters corresponding to the current user. Based on the upgrade decision parameters, differentiated upgrade instructions are sent through a remote upgrade server. For the upgraded set-top boxes, digital content and value-added applications are issued in combination with the user portraits. In the big data analysis platform, all data related to the audience rating are integrated and a global knowledge graph is constructed. The entities and relationships in the global knowledge graph are embedded into a low-dimensional semantic space through knowledge representation learning technology, a semantic association network is generated, and the similarities between different user nodes, different content nodes and different advertising nodes are calculated through a collaborative filtering algorithm, a multi-dimensional semantic matching result is generated, and a personalized content recommendation list is determined. Content is delivered according to the personalized content recommendation list, and after the delivery is completed, the effect data per unit time is recorded and the effect data is processed by an online multi-armed bandit algorithm to generate a content governance report, and the content governance report and the optimal traceability and combat plan are integrated to obtain a dual governance report.
2. The method according to claim 1, characterized in that An illegal set-top box recognition model is built based on the end-cloud collaborative architecture. Communication data packets are obtained and multi-dimensional features are extracted. The key multi-dimensional features are screened with the attention mechanism, and the interactive semantic features are obtained by fusing the key multi-dimensional features. The convolutional neural network is used for joint modeling and deep cross-learning to obtain a comprehensive feature map. The illegal behavior classifier is built through small sample learning technology and the illegal behaviors are identified, including: An illegal set-top box identification model is constructed based on a preset end-cloud collaborative architecture, a communication data packet is obtained as input data of the illegal set-top box identification model, link layer attributes, network layer attributes, transport layer attributes, application layer attributes and load content in the communication data packet are extracted, and multi-dimensional heterogeneous data are obtained by combining them. For the multi-dimensional heterogeneous data, frequency features, periodic features and object features corresponding to the communication mode are extracted by statistical analysis methods, and communication mode statistical features are obtained by combining them. For the load content, context features are extracted by natural language processing and sequence mining technology, and combined features corresponding to abnormal behaviors are extracted by combining frequent pattern mining technology. The extracted features are uploaded to the cloud, and pairwise similarity calculations are performed on all feature dimensions to generate a global feature correlation graph. The weights of the nodes and edges in the global feature correlation graph are learned through the attention mechanism to obtain the significance weight of each node and the importance weight of each edge, the global feature correlation graph is pruned and the pruned node features are extracted to obtain the key multidimensional features, the key multidimensional features are mapped to a low-dimensional vector space through a topological association algorithm based on graph embedding, the distances of different key multidimensional features in the low-dimensional vector space are calculated and the low-dimensional space similarity is determined, the temporal features are determined in combination with the behavioral events corresponding to the key multidimensional features, and the temporal causal dependencies between different key multidimensional features are calculated through a causal inference method, the kernel representation of the low-dimensional space similarity and the temporal causal dependencies is learned through a multi-core learning method, and mapped to a common semantic concept space through a multi-view joint dimensionality reduction strategy to generate the interactive semantic features; Based on the interactive semantic features and the key multidimensional features, a convolutional neural network is used for joint modeling to generate an initial comprehensive feature map, the interactive semantic features and the key multidimensional features are used as inputs of the convolutional neural network, the map embedding of the initial comprehensive feature map is used as the convolution kernel, the activation value is determined by a convolution operation, the map embedding is updated by end-to-end back propagation to obtain the comprehensive feature map, based on the node embedding vector corresponding to each node in the comprehensive feature map, a training set is constructed and a small sample classifier is set as an illegal behavior classifier, the hyperparameters of the illegal behavior classifier are randomly initialized and the training set is added to the illegal behavior classifier for training, the hyperparameters of the illegal behavior classifier are repeatedly adjusted until the illegal behavior classifier converges, the trained illegal behavior classifier is used to identify the illegal behavior of the non-intelligent set-top box, and if illegal behavior exists, the current non-intelligent set-top box is marked as an illegal set-top box.
3. The method according to claim 1, characterized in that For illegal set-top boxes, the optimal penalty action sequence is generated based on the reinforcement learning algorithm and automatically triggered through the smart contract program. The user behavior history and social relationships are learned through graph embedding representation through the graph neural network to determine the propagation path and impact chain of illegal behavior, identify key propagation nodes and simulate the effect of cutting off the propagation path based on the causal intervention algorithm, and generate the optimal tracing and crackdown plan, including: For the illegal set-top box, the penalty process is modeled as a Markov decision process through a reinforcement learning algorithm, a state space is set and state parameters are determined, the state parameters include the dynamic structure of the propagation network, the attributes of each node and the global attributes, an action space is set and the penalty intensity and resource consumption corresponding to each action are determined, at each time step, the punished node and the penalty action are selected as a combined action, a state transfer function and a reward function are defined, an action value function is set based on a deep reinforcement learning algorithm and an action with a maximum function value is selected, a value function is fitted through a deep neural network, and the optimal penalty action corresponding to the current time step is solved and obtained, the solution is repeated to obtain the optimal penalty action sequence, the optimal penalty action sequence is written into a smart contract, and the penalty execution is controlled through a contract program; Determine the current violating user based on the optimal penalty action sequence, initialize the user set and the social network of the current violating user according to the comprehensive feature map, obtain the behavior history of the current violating user, perform representation learning on the current violating user based on the graph neural network, aggregate the node information of the current violating user and the node information corresponding to the neighbor nodes of the current violating user in the social network through the graph convolution layer, aggregate the behavior history through the attention mechanism, calculate the corresponding attention weight for each behavior and perform node aggregation to obtain the user embedding representation; Based on the user embedding representation, randomly select a user pair and calculate the user similarity of the current user pair, calculate the corresponding illegal behavior propagation probability according to the user similarity, generate a propagation probability matrix and perform pruning to obtain a directed weighted propagation network, based on the directed weighted propagation network, through a depth-first search algorithm, starting from the node corresponding to the current illegal user, traverse the outgoing edge with the highest propagation probability until reaching the visited node or leaf node, obtain the illegal behavior propagation path, add randomness to the depth-first search algorithm and search again, obtain multiple propagation paths and perform a union operation on the nodes activated on each propagation path to obtain an influence chain, and determine the key propagation node based on the influence chain and the illegal behavior propagation path; A causal graph is constructed based on the causal intervention algorithm, the conditional probability of combating violations is calculated based on the causal graph and the backdoor criterion, the average causal effect is calculated based on the conditional distribution and the key propagation node with the maximum causal effect is selected based on the greedy strategy, the effect of cutting off the propagation path is simulated on the key propagation node and the optimal tracing and combat plan is generated.
4. The method according to claim 3, characterized in that The reward function is shown in the following formula: ; in, r ( s , v , a ) indicates the status s Midpoint v In Action a The return value after α 1 represents the change weight, △ N ( s , v , a ) indicates the status s Midpoint v In Action a The change value after α 2 represents the node importance weight, I ( v ) represents a node v The importance of β represents the cost weight, C ( a ) indicates an action a The corresponding cost.
5. The method according to claim 1, characterized in that The viewing behavior data of non-intelligent set-top boxes is obtained through a big data analysis platform. Based on user portrait technology, the factor decomposition machine and the attention mechanism are combined to extract implicit feature factors and construct user portraits. Based on the user portraits, the viewing preferences and interactive capabilities of each user are predicted in combination with a machine learning algorithm to obtain the upgrade decision parameters corresponding to the current user. Based on the upgrade decision parameters, differentiated upgrade instructions are sent through a remote upgrade server. For the upgraded set-top boxes, digital content and value-added applications are sent in combination with the user portraits, including: Based on a pre-set big data analysis platform, the viewing behavior data of non-intelligent set-top boxes are collected and stored, wherein the viewing behavior data includes channel switching, viewing duration and viewing time period, the viewing behavior data, users and programs are represented as triples, the viewing behavior data is modeled by a factor decomposition machine, each user and program is mapped to a low-dimensional latent vector space, the user latent vector and the program latent vector are obtained and the inner product is calculated, the current user's preference for the current program is obtained according to the calculation result, the preference degree is used as a feature interaction term, and the prediction function corresponding to the factor decomposition machine is constructed in combination with the global bias term and the feature weight; For the feature interaction term in the prediction function, the attention mechanism is used to output a scalar score value with the hidden state vector, user hidden vector and program hidden vector at the current moment as data, and the scalar score value corresponding to each time step is indexed and normalized to obtain the feature interaction term weight corresponding to each time step, and the feature interaction term weight is added to the prediction function to solve and obtain the user portrait; Constructing a user feature vector based on the user portrait and the user latent vector, training a preset user classification model through a machine learning algorithm, taking the user feature vector as input, predicting the viewing preference category and interaction ability level of the current user, and obtaining an upgrade decision parameter corresponding to the current user based on the viewing preference category and the interaction ability level; Based on the upgrade decision parameters, a differentiated upgrade instruction is sent to the target set-top box through a remote upgrade server to trigger software upgrade and function update. The differentiated upgrade instruction includes the target software version, upgrade time and upgrade priority. For the upgraded set-top box, combined with the user portrait and the upgrade decision parameters, personalized digital content and value-added applications are matched for the current user through a content filtering algorithm.
6. The method according to claim 5, characterized in that The matching formula through content filtering algorithm is as follows: ; Among them, S represents the similarity, I un Indicates user u and users n A collection of programs that have been interacted with together. f () represents the time decay function, t ui Indicates user u About the show i The timestamp of the interaction. t ni Indicates user n About the show i The timestamp of the interaction. r ui Indicates user u About the show i The preference value of r ni Indicates user n About the show i The preference value of Indicates user u Average preference for the program, Indicates user n Average preference for the program.
7. The method according to claim 1, characterized in that In the big data analysis platform, all the data related to the audience rating are integrated and a global knowledge graph is constructed. The entities and relationships in the global knowledge graph are embedded into a low-dimensional semantic space through knowledge representation learning technology, a semantic association network is generated, and similarities between different user nodes, different content nodes, and different advertising nodes are respectively calculated through a collaborative filtering algorithm, a multi-semantic matching result is generated, and a personalized content recommendation list is determined. Content is delivered according to the personalized content recommendation list, and after delivery, the effect data per unit time is recorded and the effect data is processed by an online multi-armed bandit algorithm to generate a content governance report. The content governance report and the optimal traceability strike plan are integrated to obtain a dual governance report including: Integrate all the data related to the viewing in the big data analysis platform, including users, content, interactions, advertisements and ratings, construct the global knowledge graph based on the integrated data, and for the global knowledge graph, use the relationship between each entity as a translation vector through knowledge representation learning technology, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space to obtain entity semantic vectors, and for the entity semantic vectors, calculate the semantic relevance between different entities through a preset vector space model to generate a semantic association matrix, and construct a semantic association network based on the semantic association matrix, wherein the nodes in the semantic association network represent entities, and the weighted edges between nodes represent the strength of semantic association between entities; Based on the semantic association network, combined with the collaborative filtering idea, the similarity between different users is calculated based on cosine similarity and a user similarity matrix is generated, the similarity between the target content and other content is calculated and a content similarity matrix is generated, the similarity between the target advertising node and the candidate advertising node is calculated and an advertising similarity matrix is generated, based on the user similarity matrix, the content similarity matrix and the advertising similarity matrix obtained by calculation, the comprehensive matching degree is calculated by weighted fusion to generate the multivariate semantic matching result, based on the multivariate semantic matching result, a mixed integer programming model is constructed in combination with pre-set diversity constraints and novelty constraints, the mixed integer programming model is solved by a branch and bound algorithm to obtain the personalized content recommendation list, and content delivery is performed according to the personalized content recommendation list; Record the effect data per unit time after the delivery is completed and define key performance indicators, calculate the key performance indicators of the personalized content recommendation list based on the effect data, evaluate the recommendation effect based on the key performance indicators, model the content recommendation problem as a multi-armed slot machine problem, use the recommendation strategy corresponding to the personalized content recommendation list as the slot arm, modify the selection probability of each slot arm through an online learning algorithm and improve the recommendation effect, repeatedly optimize to obtain the optimal content governance decision, generate a content governance report based on the optimal content governance decision, and integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
8. A non-intelligent set-top box dual-governance operating system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to build an illegal set-top box identification model based on the end-cloud collaborative architecture, obtain communication data packets and extract multi-dimensional features, combine the attention mechanism to perform feature screening to obtain key multi-dimensional features, fuse the key multi-dimensional features to obtain interactive semantic features, perform joint modeling and deep cross-learning through convolutional neural networks, obtain a comprehensive feature map, build an illegal behavior classifier and identify illegal behaviors through small sample learning technology, generate the optimal penalty action sequence for illegal set-top boxes based on the reinforcement learning algorithm and automatically trigger it through the smart contract program, and use the graph neural network to perform graph embedding representation learning on user behavior history and social relationships, determine the illegal behavior propagation path and impact chain, identify key propagation nodes, and simulate the effect of cutting off the propagation path based on the causal intervention algorithm to generate the optimal tracing and crackdown plan; The second unit is used to obtain the viewing behavior data of non-intelligent set-top boxes through a big data analysis platform, extract implicit feature factors based on user portrait technology, combine factor decomposition machine and attention mechanism, and construct user portraits, and predict the viewing preference and interaction ability of each user based on the user portraits in combination with machine learning algorithms to obtain the upgrade decision parameters corresponding to the current user, and send differentiated upgrade instructions through a remote upgrade server based on the upgrade decision parameters, and for the upgraded set-top boxes, send digital content and value-added applications in combination with the user portraits; The third unit is used to integrate all the data related to the viewing in the big data analysis platform and construct a global knowledge graph, embed the entities and relationships in the global knowledge graph into a low-dimensional semantic space through knowledge representation learning technology, generate a semantic association network, and calculate the similarities between different user nodes, different content nodes and different advertising nodes through a collaborative filtering algorithm, generate multi-dimensional semantic matching results and determine a personalized content recommendation list, deliver content according to the personalized content recommendation list, record the effect data per unit time after the delivery is completed, and process the effect data through an online multi-armed bandit algorithm to generate a content governance report, and integrate the content governance report and the optimal traceability and combat plan to obtain a dual governance report.
9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.
10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.
Citation Information
Patent Citations
Two-way set top box intrusion detection system and detection method thereof
CN104954864A
Knowledge graph recommendation system fusing entity information and popularity
CN112966091A
Graph neural network recommendation method and system based on commodity knowledge graph
CN114881742A
Network security information mapping system and method based on hierarchical perception
CN118300809A
Management and control method and system for intelligent viewing terminal equipment
CN118714408A
Cited By
Private domain live broadcast content distribution and visitor interaction method and system based on edge computing
CN119211585A
Model optimization method and system based on large and small model collaborative intelligence
CN119476445A
Model optimization method and system based on collaborative intelligence of large and small models
CN119476445B
Security operation and maintenance management method and system based on network security event
CN119835080A
Video multi-level adaptive compression cross-layer transmission optimization method and system based on deep learning
CN120111246A
