Private domain live broadcast content distribution and visitor interaction method and system based on edge computing

Abstract

The present invention provides a method and system for private domain live broadcast content distribution and visitor interaction based on edge computing, which relates to the field of edge computing technology, including building a distributed collaborative network by hierarchically deploying core layer, regional layer and access layer edge computing nodes; regional layer nodes collect load status information, use federated learning, LSTM time series prediction and attention mechanism to predict access popularity, build an adaptive content distribution model, and optimize content caching and distribution; access layer nodes collect user multimodal interaction data, including user input, device perception and social interaction data, identify interaction patterns through attention mechanism and integrate historical data to generate intelligent interaction strategies, so as to realize personalized recommendation and real-time interaction.

Description

Technical Field

[0001] The present invention relates to edge computing technology, and in particular to a method and system for private domain live broadcast content distribution and visitor interaction based on edge computing. Background Art

[0002] The existing technology mainly has the following defects and shortcomings:

[0003] Low content distribution efficiency: Traditional live content distribution methods usually rely on central servers. When the user scale is large or geographically dispersed, network congestion is prone to occur, resulting in slow content loading and freezing, which affects user experience. Although edge computing can improve content distribution efficiency, most existing solutions lack intelligent content caching and distribution strategies, and cannot dynamically adjust content distribution according to user access patterns and network conditions, resulting in low resource utilization.

[0004] Poor interactive experience: Existing live interactive methods are mainly text barrage and simple likes, lacking richer interactive forms and personalized interactive experience. Users' multimodal interactive data, such as voice, expression, action, etc., are not fully utilized, making it difficult to achieve real-time interaction and personalized recommendations in the true sense.

[0005] Lack of effective load balancing mechanism: In the edge computing environment, the computing and storage resources of each edge node are limited, and the user access pattern is also highly volatile. The existing load balancing mechanism is often not flexible enough to effectively respond to sudden traffic and dynamically changing user needs, which can easily cause some edge nodes to be overloaded, affecting the stability and performance of the overall system. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for private domain live broadcast content distribution and visitor interaction based on edge computing, which can solve the problems in the prior art.

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

[0008] Provides edge computing-based private domain live broadcast content distribution and visitor interaction methods, including:

[0009] Receive a live content distribution request, generate a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy includes core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establish a distributed collaborative network based on the hierarchical deployment strategy, and send node deployment and resource configuration instructions to corresponding edge computing nodes;

[0010] The regional layer nodes collect the load status information of the edge computing nodes through the distributed collaborative network, fuse the load status information of the edge computing nodes using the federated learning method, and obtain the access popularity prediction value of the live content by combining the LSTM time series prediction and the attention mechanism to build an adaptive content distribution model; generate a content caching strategy based on the adaptive content distribution model, determine the distribution location of the content on the edge node according to the content caching strategy, and generate a corresponding content distribution routing solution;

[0011] The access layer node collects multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; performs feature extraction on the multimodal interaction data, identifies the interaction mode through the attention mechanism, and integrates the historical interaction data to generate an intelligent interaction strategy, and the intelligent interaction strategy includes a personalized recommendation strategy and a real-time interaction plan, and sends the intelligent interaction strategy to the corresponding edge computing node.

[0012] In an optional embodiment,

[0013] Receiving a live content distribution request, generating a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy including core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establishing a distributed collaborative network based on the hierarchical deployment strategy, and sending node deployment and resource configuration instructions to corresponding edge computing nodes include:

[0014] Receive a live content distribution request, the live content distribution request including an anchor identifier, a live broadcast type, and information about an expected audience size; generate a hierarchical deployment strategy using a hierarchical clustering algorithm according to a geographic location, a network topology, and computing resources;

[0015] Based on the hierarchical deployment strategy, a two-phase commit protocol is adopted to establish a distributed collaborative network, including: sending a pre-deployment message containing node deployment and resource configuration to all layer nodes in the preparation phase, and issuing a deployment instruction after receiving confirmation information from all layer nodes in the commit phase;

[0016] The functions of nodes at each layer are configured based on the deployment instructions, including: the core layer nodes establish a constraint optimization model for resource allocation based on a dynamic programming algorithm, and send the optimized resource allocation results to the regional layer nodes; the regional layer nodes use a minimum spanning tree algorithm based on the optimized resource allocation results to build an edge content distribution tree, and the edge weights of the edge content distribution tree are obtained by weighted calculation of network bandwidth, delay and jitter; the access layer nodes implement load balancing distribution of user requests based on a consistent hashing algorithm and a virtual node mechanism.

[0017] In an optional embodiment,

[0018] The core layer node establishes a constrained optimization model for resource allocation based on a dynamic programming algorithm, and sends the optimized resource allocation result to the regional layer node; the regional layer node uses a minimum spanning tree algorithm to construct an edge content distribution tree based on the optimized resource allocation result, and the edge weight of the edge content distribution tree is obtained by weighted calculation of network bandwidth, delay and jitter; the access layer node implements the load balancing distribution of user requests based on a consistent hashing algorithm and a virtual node mechanism, including the following steps:

[0019] A resource allocation constraint optimization model for core layer nodes is constructed based on a dynamic programming algorithm and a Lagrange multiplier method. The resource allocation constraint optimization model includes allocation parameters for computing resources, storage resources, and bandwidth resources. An optimized resource allocation plan is obtained by solving an extreme value equation under resource allocation constraint conditions. The optimized resource allocation plan is sent to regional layer nodes.

[0020] The regional layer node constructs an edge content distribution tree based on the optimized resource allocation scheme, and establishes an edge content distribution network topology using a minimum spanning tree algorithm. The edge weights of the edge content distribution tree are obtained by adaptive weighted calculation of bandwidth, delay and jitter values. The weight coefficients of the adaptive weighting are dynamically adjusted based on a distribution performance metric function using an online learning method, and the topological structure information of the edge content distribution tree is sent down to the access layer node;

[0021] The access layer node uses a consistent hashing algorithm and a dynamic virtual node adjustment mechanism based on the topology information to achieve load balancing distribution of user requests. The dynamic virtual node adjustment mechanism adaptively adjusts the number of virtual nodes according to the actual load factor of the node, and adjusts the distribution ratio of user requests among the nodes by dynamically adjusting the virtual node distribution;

[0022] A multi-level feedback adjustment mechanism is established, wherein the access layer nodes periodically collect load status information and feed it back to the regional layer nodes for adjusting the edge content distribution tree, and the regional layer nodes collect content distribution efficiency indicators and feed them back to the core layer nodes for optimizing resource allocation solutions.

[0023] In an optional embodiment,

[0024] The regional layer nodes collect the load status information of the edge computing nodes through the distributed collaborative network, fuse the load status information of the edge computing nodes using the federated learning method, and obtain the access popularity prediction value of the live content by combining the LSTM time series prediction and the attention mechanism. The steps of constructing the adaptive content distribution model include:

[0025] The load status information includes computing load information, storage capacity information and network bandwidth information;

[0026] A load feature extraction model is maintained locally at each edge computing node, the input of the load feature extraction model is the state sequence of the node within a preset time window, and the output is the load feature vector of the node; the load feature vector is input into the predictor to obtain a predicted load value, and a fusion loss function is constructed based on the mean square error between the predicted load value and the actual load and the cosine similarity between the current feature and the historical feature; each edge computing node independently trains a local load feature extraction model based on the fusion loss function, and regularly uploads the load feature extraction model parameters to the coordination node; the coordination node determines the weight according to the computing power and data quality of the edge computing node, and distributes the received load feature extraction model parameters to each edge computing node after weighted aggregation;

[0027] The load feature vector is concatenated with the historical access record to construct an input sequence, and a bidirectional long short-term memory network is used to extract time series features; an attention mechanism is introduced to calculate the importance weights of different time steps, and the time series features are weighted and summed based on the importance weights to obtain a context vector; the context vector is processed through a fully connected layer to obtain an access heat prediction value;

[0028] Based on the load feature vector and access popularity prediction value of the edge computing node, a multi-objective optimization method is used to weightedly combine the load balancing target, access delay target and bandwidth cost target to construct a comprehensive objective function; wherein the load balancing target is used to minimize the maximum value of the node load and capacity ratio, the access delay target is used to minimize the sum of delays for users to access the node, and the bandwidth cost target is used to minimize the sum of bandwidth costs for content transmission between nodes; and an adaptive content distribution model is obtained by solving the comprehensive objective function.

[0029] In an optional embodiment,

[0030] The steps of generating a content caching strategy based on the adaptive content distribution model, determining the distribution location of the content at the edge node according to the content caching strategy, and generating a corresponding content distribution routing solution include:

[0031] Generate a content caching strategy based on an adaptive content distribution model, wherein the content caching strategy is determined by calculating the matching degree between the content and the edge computing node, wherein the matching degree is calculated based on the load characteristic vector, processing capacity, and access popularity prediction value of the edge computing node, and when the matching degree exceeds a preset threshold, the corresponding content is allocated to the corresponding edge computing node for caching;

[0032] Constructing a content node allocation matrix according to the content caching strategy; determining the distribution location of the content on the edge computing node based on the content node allocation matrix, executing a single allocation constraint on each content, and executing a capacity constraint on each edge computing node;

[0033] A reachability graph is constructed based on the distribution position of the content in the edge computing nodes, wherein the reachability graph includes a set of edge computing nodes and connection relationships between the nodes; in response to a user content access request, a set of source nodes that cache the access request content and edge computing nodes that the user can directly access are determined from the content node allocation matrix, and a path cost from the edge computing node that the user can directly access to each node in the source node set is calculated, wherein the path cost is calculated based on the number of path hops and the average load rate of the nodes on the path; and a path with the minimum path cost is selected as a content distribution routing solution.

[0034] In an optional embodiment,

[0035] The access layer node collects multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extracts features from the multimodal interaction data, identifies interaction patterns through an attention mechanism, and integrates historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions. The steps of sending the intelligent interaction strategies to corresponding edge computing nodes include:

[0036] Based on the bidirectional long short-term memory network, the user input data is subjected to temporal feature extraction to obtain a user input feature vector, based on the compressed convolutional network, the device perception data is subjected to spatial feature extraction to obtain a device perception feature vector, and based on the graph attention network, the social interaction data is subjected to interaction feature extraction to obtain a social interaction feature vector;

[0037] The user input feature vector, the device perception feature vector and the social interaction feature vector are spliced ​​to form a multimodal feature, the feature correlation of the multimodal feature is calculated through a multi-head attention mechanism, and an interaction pattern vector is generated based on the feature correlation; the interaction pattern vector is fused with the historical interaction sequence by using a gated update mechanism to obtain a fused interaction state;

[0038] A dual-branch strategy network is constructed based on the fused interaction state, wherein the first branch of the dual-branch strategy network is used to generate a personalized recommendation strategy, and the second branch of the dual-branch strategy network is used to generate a real-time interaction plan. The personalized recommendation strategy and the real-time interaction plan are strategically combined to obtain an intelligent interaction strategy.

[0039] In an optional embodiment,

[0040] A dual-branch strategy network is constructed based on the fused interaction state, wherein the first branch of the dual-branch strategy network is used to generate a personalized recommendation strategy, and the second branch of the dual-branch strategy network is used to generate a real-time interaction plan. The steps of strategically combining the personalized recommendation strategy and the real-time interaction plan to obtain an intelligent interaction strategy include:

[0041] A shared coding layer is constructed by using a multi-layer perceptron, and the fused interaction state is input into the shared coding layer to obtain coding features, and the coding features are used for two parallel branches of the dual-branch strategy network;

[0042] The first branch of the dual-branch strategy network uses an Actor-Critic architecture to generate a personalized recommendation strategy; wherein the Actor network includes a multi-layer strategy neural network, generates an Actor hidden state based on the encoding feature, and outputs a recommended action probability distribution as a preliminary personalized recommendation strategy through a softmax activation function; the Critic network includes a multi-layer value neural network, generates a Critic hidden state based on the encoding feature and outputs a state value estimate, calculates a temporal difference error using a temporal difference learning method, updates the Actor network parameters and the Critic network parameters based on the temporal difference error, introduces an entropy regularization term to construct an Actor network loss function to optimize the preliminary personalized recommendation strategy to obtain a final personalized recommendation strategy;

[0043] The second branch of the dual-branch strategy network uses a conditional variational autoencoder to generate a real-time interaction plan, including inputting the encoding feature into an encoder to obtain a mean vector and a variance vector as latent variable distribution parameters, sampling latent variables based on the latent variable distribution parameters through a reparameterization method, splicing the latent variables with the encoding feature, and then sequentially performing time series modeling through a decoding multi-layer perceptron and a gated recurrent unit, and inputting the hidden state of the gated recurrent unit into a generator network to obtain an interaction instruction to form a real-time interaction plan;

[0044] Construct a strategy combination module, the strategy combination module includes: a mutual information calculation unit, used to calculate the mutual information between the personalized recommendation strategy and the real-time interaction plan; a joint optimization unit, used to construct a joint optimization target using the mutual information as a regularization term; a strategy fusion unit, used to calculate the interaction strategy weight coefficient based on the joint optimization target, perform weighted combination of the personalized recommendation strategy and the real-time interaction plan, and generate an intelligent interaction strategy.

[0045] According to a second aspect of the embodiments of the present invention,

[0046] Provides a private domain live content distribution and visitor interaction system based on edge computing, including:

[0047] The first unit is used to receive a live content distribution request, generate a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy includes core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establish a distributed collaborative network based on the hierarchical deployment strategy, and send node deployment and resource configuration instructions to corresponding edge computing nodes;

[0048] The second unit is used for the regional layer nodes to collect the load status information of the edge computing nodes through the distributed collaborative network, adopt the federated learning method to fuse the load status information of the edge computing nodes, combine the LSTM time series prediction and attention mechanism to obtain the access popularity prediction value of the live content, and build an adaptive content distribution model; generate a content caching strategy based on the adaptive content distribution model, determine the distribution position of the content on the edge node according to the content caching strategy, and generate a corresponding content distribution routing solution;

[0049] The third unit is used for the access layer node to collect multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extract features from the multimodal interaction data, identify interaction patterns through an attention mechanism, and integrate historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions, and send the intelligent interaction strategies to corresponding edge computing nodes.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of the embodiments of the present invention,

[0056] 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.

[0057] The present invention can improve the efficiency of live content distribution: by hierarchically deploying edge computing nodes and an adaptive content distribution model, accurate content caching and distribution can be performed according to the predicted access popularity of the live content, thereby reducing network latency and improving user access speed and quality of experience.

[0058] The present invention can enhance the user interaction experience: by utilizing multimodal interaction data and attention mechanism, it can identify the user's interaction pattern, and generate personalized recommendation strategies and real-time interaction solutions in combination with historical data, thereby enhancing user participation and interaction experience.

[0059] The present invention can optimize resource utilization: by fusing the load status information of edge computing nodes based on the federated learning method, dynamic scheduling and optimization of global resources can be achieved, resource utilization can be improved, and operating costs can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flowchart of a method for private domain live broadcast content distribution and visitor interaction based on edge computing according to an embodiment of the present invention;

[0061] Figure 2 This is a structural diagram of a private domain live content distribution and visitor interaction system based on edge computing in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] 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.

[0063] 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.

[0064] Figure 1 Schematic diagram of the process of the method for private domain live broadcast content distribution and visitor interaction based on edge computing according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0065] S1. Receive a live content distribution request, generate a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy includes core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establish a distributed collaborative network based on the hierarchical deployment strategy, and send node deployment and resource configuration instructions to the corresponding edge computing nodes;

[0066] S2. The regional layer nodes collect the load status information of the edge computing nodes through the distributed collaborative network, fuse the load status information of the edge computing nodes using the federated learning method, and obtain the access popularity prediction value of the live content by combining the LSTM time series prediction and attention mechanism to build an adaptive content distribution model; generate a content caching strategy based on the adaptive content distribution model, determine the distribution location of the content on the edge node according to the content caching strategy, and generate a corresponding content distribution routing solution;

[0067] S3. The access layer node collects multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; performs feature extraction on the multimodal interaction data, identifies interaction patterns through an attention mechanism, and integrates historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions, and sends the intelligent interaction strategies to corresponding edge computing nodes.

[0068] In an optional embodiment,

[0069] Receiving a live content distribution request, generating a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy including core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establishing a distributed collaborative network based on the hierarchical deployment strategy, and sending node deployment and resource configuration instructions to corresponding edge computing nodes include:

[0070] Receive a live content distribution request, the live content distribution request including an anchor identifier, a live broadcast type, and information about an expected audience size; generate a hierarchical deployment strategy using a hierarchical clustering algorithm according to a geographic location, a network topology, and computing resources;

[0071] Based on the hierarchical deployment strategy, a two-phase commit protocol is adopted to establish a distributed collaborative network, including: sending a pre-deployment message containing node deployment and resource configuration to all layer nodes in the preparation phase, and issuing a deployment instruction after receiving confirmation information from all layer nodes in the commit phase;

[0072] The functions of nodes at each layer are configured based on the deployment instructions, including: the core layer nodes establish a constraint optimization model for resource allocation based on a dynamic programming algorithm, and send the optimized resource allocation results to the regional layer nodes; the regional layer nodes use a minimum spanning tree algorithm based on the optimized resource allocation results to build an edge content distribution tree, and the edge weights of the edge content distribution tree are obtained by weighted calculation of network bandwidth, delay and jitter; the access layer nodes implement load balancing distribution of user requests based on a consistent hashing algorithm and a virtual node mechanism.

[0073] Exemplarily, first, the system receives a live content distribution request, which includes key information, such as the anchor ID (e.g., the anchor ID is "12345"), the live broadcast type (e.g., game live broadcast, music live broadcast), and the expected audience size (e.g., 10,000 people).

[0074] Next, the system generates a hierarchical deployment strategy based on geographic location, network topology, and computing resources. For example, suppose there are three regions of audiences: Region A (5,000 people), Region B (3,000 people), and Region C (2,000 people). The system uses a hierarchical clustering algorithm to divide edge computing nodes into core layer, regional layer, and access layer. The core layer nodes are responsible for global resource scheduling, the regional layer nodes are responsible for regional content distribution optimization, and the access layer nodes are responsible for user interaction. Assume that the core layer nodes are located in the geographic center, the regional layer nodes are located in the center of regions A, B, and C, and the access layer nodes are distributed in each region.

[0075] Then, the system establishes a distributed collaborative network. A two-phase commit protocol is used to ensure that all nodes are deployed synchronously. In the preparation phase, the system sends a pre-deployment message to all layer nodes, which contains node deployment and resource configuration information. For example, 5 virtual machines are allocated to the regional layer nodes of region A, and each virtual machine is configured with an 8-core CPU and 16GB of memory. In the commit phase, after receiving confirmation information from all layer nodes, the system issues deployment instructions. If a node is not confirmed, the deployment process is rolled back.

[0076] Subsequently, the system configures the functions of nodes at each layer. The core layer nodes establish a constrained optimization model for resource allocation based on a dynamic programming algorithm. For example, the core layer nodes determine the resources required for each regional layer node based on the expected audience size and regional distribution, for example, 60% of computing resources are allocated to region A, 30% to region B, and 10% to region C. The optimized resource allocation results are sent to the regional layer nodes. Based on the received resource allocation results, the regional layer nodes use the minimum spanning tree algorithm to build an edge content distribution tree. For example, the regional layer node in region A selects the best path to distribute the live content to its subordinate access layer nodes based on network bandwidth, latency, and jitter indicators. The edge weight is calculated by weighting these indicators, for example, the bandwidth weight is 0.5, the latency weight is 0.3, and the jitter weight is 0.2. The access layer nodes implement load balancing distribution of user requests based on the consistent hashing algorithm and the virtual node mechanism. For example, 100 virtual nodes are used to evenly distribute user requests to different access layer nodes to ensure load balancing of each node.

[0077] The present invention reduces the data transmission distance by caching the content to the edge node close to the user, thereby reducing the live broadcast delay and improving the user experience; the resource allocation and content distribution path optimization are performed through the dynamic programming algorithm and the minimum spanning tree algorithm, which maximizes the use of edge computing resources and avoids resource waste; the two-phase commit protocol and the consistent hashing algorithm ensure the stability and reliability of the system, and the system can still operate normally even when some nodes fail.

[0078] In an optional embodiment,

[0079] The core layer node establishes a constrained optimization model for resource allocation based on a dynamic programming algorithm, and sends the optimized resource allocation result to the regional layer node; the regional layer node uses a minimum spanning tree algorithm to construct an edge content distribution tree based on the optimized resource allocation result, and the edge weight of the edge content distribution tree is obtained by weighted calculation of network bandwidth, delay and jitter; the access layer node implements the load balancing distribution of user requests based on a consistent hashing algorithm and a virtual node mechanism, including the following steps:

[0080] A resource allocation constraint optimization model for core layer nodes is constructed based on a dynamic programming algorithm and a Lagrange multiplier method. The resource allocation constraint optimization model includes allocation parameters for computing resources, storage resources, and bandwidth resources. An optimized resource allocation plan is obtained by solving an extreme value equation under resource allocation constraint conditions. The optimized resource allocation plan is sent to regional layer nodes.

[0081] The regional layer node constructs an edge content distribution tree based on the optimized resource allocation scheme, and establishes an edge content distribution network topology using a minimum spanning tree algorithm. The edge weights of the edge content distribution tree are obtained by adaptive weighted calculation of bandwidth, delay and jitter values. The weight coefficients of the adaptive weighting are dynamically adjusted based on a distribution performance metric function using an online learning method, and the topological structure information of the edge content distribution tree is sent down to the access layer node;

[0082] The access layer node uses a consistent hashing algorithm and a dynamic virtual node adjustment mechanism based on the topology information to achieve load balancing distribution of user requests. The dynamic virtual node adjustment mechanism adaptively adjusts the number of virtual nodes according to the actual load factor of the node, and adjusts the distribution ratio of user requests among the nodes by dynamically adjusting the virtual node distribution;

[0083] A multi-level feedback adjustment mechanism is established, wherein the access layer nodes periodically collect load status information and feed it back to the regional layer nodes for adjusting the edge content distribution tree, and the regional layer nodes collect content distribution efficiency indicators and feed them back to the core layer nodes for optimizing resource allocation solutions.

[0084] Exemplarily, first, the core layer node collects the resource demand information of each regional layer node, including the demand for computing resources, storage resources and bandwidth resources. Then, the core layer node establishes a constrained optimization model for resource allocation based on the collected resource demand information. The goal of this model is to maximize the resource utilization of the entire system while meeting the minimum resource requirements of each regional layer node. In order to achieve this goal, a dynamic programming algorithm is used to decompose the resource allocation problem into multiple sub-problems and solve them layer by layer. For example, the entire resource pool can be divided into multiple resource blocks, and then the resource blocks are gradually allocated according to the requirements of each regional layer node. In order to avoid uneven resource allocation, the Lagrange multiplier method is introduced to convert the resource constraints into part of the objective function, thereby converting the constrained optimization problem into an unconstrained optimization problem. Through iterative solution, the optimized resource allocation scheme is finally obtained, including the number of computing resources, storage resources and bandwidth resources allocated to each regional layer node. For example, the core layer node has a total of 1,000 computing units, 500TB storage space and 10Gbps bandwidth. Regional node A requests 200 computing units, 100TB storage space, and 2Gbps bandwidth, and regional node B requests 300 computing units, 200TB storage space, and 3Gbps bandwidth. After optimization, the core node allocates 250 computing units, 120TB storage space, and 2.5Gbps bandwidth to regional node A, and 350 computing units, 220TB storage space, and 3.5Gbps bandwidth to regional node B, leaving the remaining resources for subsequent allocation. Finally, the core node sends the optimized resource allocation plan to the corresponding regional node.

[0085] After receiving the resource allocation plan from the core layer node, the regional layer node starts to build an edge content distribution tree. The goal of this distribution tree is to efficiently distribute content to each access layer node. In order to build a minimum spanning tree, it is necessary to first determine the nodes and edges. The nodes represent the access layer nodes, and the edges represent the connections between the nodes. The weight of the edge is calculated by weighting the network bandwidth, delay, and jitter. For example, the weight of a link with a bandwidth of 100Mbps, a delay of 20ms, and a jitter of 5ms can be calculated as 0.5*100+0.3*(1 / 20)+0.2*(1 / 5). Among them, the weight coefficients 0.5, 0.3, and 0.2 represent the importance of bandwidth, delay, and jitter, respectively. These weight coefficients can be dynamically adjusted through online learning methods. The online learning method adjusts the weight coefficients in real time according to the distribution performance measurement function, such as content download speed and user experience quality, to adapt to changes in the network environment. For example, if the network bandwidth becomes a bottleneck, the weight coefficient of the bandwidth is increased. The minimum spanning tree algorithm, such as the Kruskal algorithm, is used to build the edge content distribution tree. After the construction is completed, the regional layer node sends the topological structure information of the edge content distribution tree to the access layer node. Assume that the regional layer node needs to connect to three access layer nodes, and the bandwidth, delay and jitter values ​​between the nodes are: (100Mbps, 20ms, 5ms), (50Mbps, 30ms, 10ms), (200Mbps, 10ms, 2ms). After calculation and the minimum spanning tree algorithm, the edge content distribution tree finally constructed may choose the link with the highest bandwidth and the lowest delay.

[0086] After receiving the topology information from the regional layer nodes, the access layer nodes use the consistent hashing algorithm and the virtual node mechanism to implement load balancing distribution of user requests. The consistent hashing algorithm maps the user request to a position on the hash ring, and then distributes the request to the first virtual node in the clockwise direction on the hash ring. In order to improve the accuracy of load balancing, the virtual node mechanism is introduced. Each access layer node corresponds to multiple virtual nodes, and the virtual nodes are evenly distributed on the hash ring. By dynamically adjusting the number of virtual nodes, the distribution ratio of user requests among nodes can be dynamically adjusted. For example, if the load of an access layer node is too high, reduce the number of its virtual nodes; otherwise, increase the number of its virtual nodes. The load balancing adjustment is based on the actual load factor of the node. For example, set the threshold of the load factor to 0.8. When the load factor of a node exceeds 0.8, reduce the number of its virtual nodes. Assume that there are three access layer nodes, and each node has 100 virtual nodes in the initial state. When the load factor of node A reaches 0.9, reduce the number of virtual nodes of node A to 80, and redistribute these virtual nodes to other nodes.

[0087] In order to further optimize the system performance, a multi-level feedback adjustment mechanism is established. The access layer nodes periodically collect load status information, such as CPU utilization, memory occupancy, and network traffic, and feed this information back to the regional layer nodes. The regional layer nodes adjust the topology of the edge content distribution tree based on the received load status information, such as adding or deleting nodes, adjusting link bandwidth, etc. The regional layer nodes also collect content distribution efficiency indicators, such as average download speed and content distribution success rate, and feed this information back to the core layer nodes. The core layer nodes optimize the resource allocation scheme based on the received content distribution efficiency indicators, such as adjusting the resource quota of each regional layer node. This multi-level feedback adjustment mechanism allows the system to dynamically adjust according to the actual operation situation, thereby improving the overall performance of the system. For example, if a certain access layer node continues to have too high a load, the regional layer node can adjust the distribution tree to direct part of the traffic to other nodes with lower loads. At the same time, the regional layer node will feed back this situation to the core layer node, and the core layer node can increase the resource quota of the area to fundamentally solve the problem of excessive load.

[0088] The resource allocation constraint optimization model constructed by the present invention through the dynamic programming algorithm and the Lagrange multiplier method can maximize the resource utilization and avoid resource waste; the edge content distribution tree constructed based on the minimum spanning tree algorithm, and the load balancing distribution realized by the consistent hashing algorithm and the virtual node mechanism can effectively reduce the content distribution delay and improve the content distribution efficiency; the multi-level feedback adjustment mechanism enables the system to dynamically adjust the resource allocation and content distribution strategy according to the actual operation situation, thereby enhancing the adaptability and robustness of the system.

[0089] In an optional embodiment,

[0090] The regional layer nodes collect the load status information of the edge computing nodes through the distributed collaborative network, fuse the load status information of the edge computing nodes using the federated learning method, and obtain the access popularity prediction value of the live content by combining the LSTM time series prediction and the attention mechanism. The steps of constructing the adaptive content distribution model include:

[0091] The load status information includes computing load information, storage capacity information and network bandwidth information;

[0092] A load feature extraction model is maintained locally at each edge computing node, the input of the load feature extraction model is the state sequence of the node within a preset time window, and the output is the load feature vector of the node; the load feature vector is input into the predictor to obtain a predicted load value, and a fusion loss function is constructed based on the mean square error between the predicted load value and the actual load and the cosine similarity between the current feature and the historical feature; each edge computing node independently trains a local load feature extraction model based on the fusion loss function, and regularly uploads the load feature extraction model parameters to the coordination node; the coordination node determines the weight according to the computing power and data quality of the edge computing node, and distributes the received load feature extraction model parameters to each edge computing node after weighted aggregation;

[0093] The load feature vector is concatenated with the historical access record to construct an input sequence, and a bidirectional long short-term memory network is used to extract time series features; an attention mechanism is introduced to calculate the importance weights of different time steps, and the time series features are weighted and summed based on the importance weights to obtain a context vector; the context vector is processed through a fully connected layer to obtain an access heat prediction value;

[0094] Based on the load feature vector and access popularity prediction value of the edge computing node, a multi-objective optimization method is used to weightedly combine the load balancing target, access delay target and bandwidth cost target to construct a comprehensive objective function; wherein the load balancing target is used to minimize the maximum value of the node load and capacity ratio, the access delay target is used to minimize the sum of delays for users to access the node, and the bandwidth cost target is used to minimize the sum of bandwidth costs for content transmission between nodes; and an adaptive content distribution model is obtained by solving the comprehensive objective function.

[0095] Exemplarily, the regional layer nodes collect load status information of edge computing nodes through a distributed collaborative network, and use this information to build an adaptive content distribution model to optimize the distribution of live content.

[0096] First, define the load status information. The load status information covers three dimensions: computing load information, such as CPU utilization and memory usage; storage capacity information, such as the remaining storage space; and network bandwidth information, such as the current available bandwidth. For example, the load status information of edge computing node 1 can be expressed as: CPU utilization 70%, remaining storage space 10GB, and available bandwidth 20Mbps.

[0097] Next, a load feature extraction model is deployed on each edge computing node. The input of the model is the state sequence of the node within a preset time window (for example, the past 1 hour). For example, if the load status information is collected once a minute, the length of the input sequence is 60. The output of the model is a load feature vector that can characterize the load status of the node. For example, a simple load feature vector can be a weighted average of three load status information. To train this model, we use a predictor to predict future load values ​​based on the load feature vector. Then, a fusion loss function is constructed based on the square of the difference between the predicted load value and the actual load value and the similarity between the current feature vector and the historical feature vector. Each edge computing node independently uses this loss function to train its local load feature extraction model. For example, if the predicted CPU utilization is very different from the actual CPU utilization, or the current load feature vector is very different from the historical feature vector, the value of the loss function will be high. After training is completed, each edge computing node regularly uploads the trained model parameters to the coordination node. The coordination node assigns weights to each edge computing node based on the computing power and data quality. For example, if a node has strong computing power and high data quality, it will be given a higher weight. The coordination node performs weighted averaging on all received model parameters and then distributes the aggregated model parameters back to each edge computing node to ensure that all nodes use a consistent load feature extraction model.

[0098] Then, a visit popularity prediction model is constructed. The extracted load feature vector is concatenated with the historical visit records (such as the number of visits in the past 24 hours) to form an input sequence. The bidirectional long short-term memory network (Bi-LSTM) is used to extract the time series features of the sequence. Bi-LSTM can consider both past and future information at the same time to better capture the changing trend of visit popularity. For example, if the number of visits to a live content has continued to rise in the past few hours, then Bi-LSTM can predict that the number of visits in the future will continue to rise. The attention mechanism is introduced to calculate the importance weight of each time step. For example, if the number of visits in a certain time period surges, then the time period will be given a higher weight. Based on these weights, the time series features extracted by Bi-LSTM are weighted and summed to obtain the context vector. Finally, the context vector is processed through the fully connected layer to obtain the visit popularity prediction value.

[0099] Finally, an adaptive content distribution model is constructed. Based on the load feature vectors and access popularity prediction values ​​of edge computing nodes, a comprehensive objective function is constructed. This objective function contains three sub-objectives: a load balancing objective, which aims to minimize the maximum value of the ratio of node load to capacity to avoid excessive load on individual nodes; an access delay objective, which aims to minimize the sum of delays for users to access content and improve user experience; and a bandwidth cost objective, which aims to minimize the sum of bandwidth costs for content transmission between nodes and reduce operating costs. These three objectives form a comprehensive objective function through a weighted combination. For example, the weights of the three objectives can be adjusted according to actual needs. For example, if you pay more attention to user experience, you can increase the weight of the access delay objective. By solving this comprehensive objective function, the optimal content distribution strategy can be obtained.

[0100] The present invention can maximize the use of existing resources and avoid resource waste by balancing the load of each edge node; by distributing content to edge nodes that are closer to users and have lower loads, it can effectively reduce the delay in user access to content and improve user experience; by optimizing the content distribution path, it can reduce the bandwidth cost of content transmission between nodes, thereby reducing operating costs.

[0101] In an optional embodiment,

[0102] The steps of generating a content caching strategy based on the adaptive content distribution model, determining the distribution location of the content at the edge node according to the content caching strategy, and generating a corresponding content distribution routing solution include:

[0103] Generate a content caching strategy based on an adaptive content distribution model, wherein the content caching strategy is determined by calculating the matching degree between the content and the edge computing node, wherein the matching degree is calculated based on the load characteristic vector, processing capacity, and access popularity prediction value of the edge computing node, and when the matching degree exceeds a preset threshold, the corresponding content is allocated to the corresponding edge computing node for caching;

[0104] Constructing a content node allocation matrix according to the content caching strategy; determining the distribution location of the content on the edge computing node based on the content node allocation matrix, executing a single allocation constraint on each content, and executing a capacity constraint on each edge computing node;

[0105] A reachability graph is constructed based on the distribution position of the content in the edge computing nodes, wherein the reachability graph includes a set of edge computing nodes and connection relationships between the nodes; in response to a user content access request, a set of source nodes that cache the access request content and edge computing nodes that the user can directly access are determined from the content node allocation matrix, and a path cost from the edge computing node that the user can directly access to each node in the source node set is calculated, wherein the path cost is calculated based on the number of path hops and the average load rate of the nodes on the path; and a path with the minimum path cost is selected as a content distribution routing solution.

[0106] Exemplarily, first, a content caching strategy is generated. The core of this strategy is to calculate the matching degree between the content and the edge computing node. The matching degree is not specified arbitrarily, but takes into account the three key factors of the load feature vector, processing capacity and access heat prediction value of the edge computing node. The load feature vector may include indicators such as CPU usage, memory occupancy, network IO, etc. For example, the load feature vector of node A is [0.7, 0.8, 0.6], indicating that its CPU usage is 70%, memory occupancy is 80%, and network IO is 60%. The processing capacity refers to the amount of content that the node can process, for example, the processing capacity of node A is 1000. The access heat prediction value is the frequency of future access to the content predicted based on historical data. For example, the access heat prediction value of content X is 500, indicating that the content is expected to be accessed 500 times in the future. Taking these factors into consideration, for example, by weighted average, the matching degree between the content and the node is calculated. When the matching degree exceeds a preset threshold, such as 0.8, the corresponding content is allocated to the corresponding edge computing node for caching. For example, if the calculated matching result between content X and node A is 0.85, which exceeds the preset threshold value 0.8, content X is cached to node A.

[0107] Next, construct the content node allocation matrix. This matrix records the allocation relationship between content and edge nodes. The rows of the matrix represent content, the columns represent edge nodes, and the values ​​of the matrix elements indicate whether the corresponding content is cached on the corresponding nodes. For example, if content X is cached on nodes A and B, the element values ​​of the rows corresponding to content X in the matrix and the columns corresponding to nodes A and B are 1, and the values ​​of the remaining elements are 0. A single allocation constraint is enforced for each content, that is, each content can only be cached on one node. A capacity constraint is enforced for each edge computing node, that is, the amount of content cached by each node cannot exceed its processing capacity. For example, if the capacity of node A is 1000, the sum of the element values ​​of the corresponding column of node A cannot exceed 1000.

[0108] Then, a reachability graph is constructed based on the content node allocation matrix. The graph describes the connection relationship between edge computing nodes. The nodes in the graph represent edge computing nodes, and the edges between nodes represent the connections between nodes. For example, if there is a connection between node A and node B, there is an edge between node A and node B in the reachability graph.

[0109] Finally, a content distribution routing scheme is generated for user content access requests. First, the source node set that caches the access request content and the edge computing nodes that the user can directly access are determined from the content node allocation matrix. For example, the user requests to access content X, content X is cached in nodes A and B, and the user can directly access nodes C and D. The source node set is {A, B}, and the node set that the user can directly access is {C, D}. Then, the path cost from the node that the user can directly access to each node in the source node set is calculated. The path cost is calculated based on the number of path hops and the average load rate of the nodes on the path. For example, the number of path hops from node C to node A is 2, and the average load rate of the nodes on the path is 0.7, then the path cost can be calculated as 2*0.7=1.4. Select the path with the smallest path cost as the content distribution routing scheme. For example, if the path cost from node C to node A is the smallest, then select the path as the content distribution routing scheme.

[0110] By caching content to edge nodes, the present invention allows users to directly obtain content from nearby edge nodes, reducing data transmission delays and thus improving content access speed; by caching content to edge nodes, the access pressure on the core network is reduced, thereby reducing network load and improving network stability; by caching and routing content according to node load, capacity and access popularity, the resources of edge nodes can be better utilized, resource waste can be avoided, and resource utilization can be improved.

[0111] In an optional embodiment,

[0112] The access layer node collects multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extracts features from the multimodal interaction data, identifies interaction patterns through an attention mechanism, and integrates historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions. The steps of sending the intelligent interaction strategies to corresponding edge computing nodes include:

[0113] Based on the bidirectional long short-term memory network, the user input data is subjected to temporal feature extraction to obtain a user input feature vector, based on the compressed convolutional network, the device perception data is subjected to spatial feature extraction to obtain a device perception feature vector, and based on the graph attention network, the social interaction data is subjected to interaction feature extraction to obtain a social interaction feature vector;

[0114] The user input feature vector, the device perception feature vector and the social interaction feature vector are spliced ​​to form a multimodal feature, the feature correlation of the multimodal feature is calculated through a multi-head attention mechanism, and an interaction pattern vector is generated based on the feature correlation; the interaction pattern vector is fused with the historical interaction sequence by using a gated update mechanism to obtain a fused interaction state;

[0115] A dual-branch strategy network is constructed based on the fused interaction state, wherein the first branch of the dual-branch strategy network is used to generate a personalized recommendation strategy, and the second branch of the dual-branch strategy network is used to generate a real-time interaction plan. The personalized recommendation strategy and the real-time interaction plan are strategically combined to obtain an intelligent interaction strategy.

[0116] Exemplarily, the access layer nodes collect multimodal interaction data of users based on the distributed collaborative network. These data cover three dimensions: user input data, device perception data, and social interaction data.

[0117] First, the user input data is processed. User input data can be in various forms such as text, voice, and image. The system uses a bidirectional long short-term memory network to extract temporal features from this data. For example, if a user continuously inputs text information such as "music", "play", and "popular songs" over a period of time, the bidirectional long short-term memory network will capture the temporal relationship between these information and convert it into a user input feature vector. Assuming that the text data input by the user is encoded as [1, 5, 3, 7], after being processed by the bidirectional long short-term memory network, the resulting feature vector may be [0.2, 0.5, 0.1, 0.8].

[0118] At the same time, the system processes the device perception data. The device perception data may include information such as the user's geographic location, device type, and ambient light. The system uses a compressed convolutional network to extract spatial features from this data. For example, the user's geographic location information can be represented as longitude and latitude coordinates. The compressed convolutional network can capture the spatial features of the user's location and convert it into a device perception feature vector. Assuming that the user's geographic location information is encoded as [2.5, 51.7], after being processed by the compressed convolutional network, the resulting feature vector may be [0.1, 0.9].

[0119] In addition, the system also processes the user's social interaction data. Social interaction data may include information about the user's interaction with other users, such as likes, comments, and reposts. The system uses a graph attention network to extract interaction features from this data. For example, a user's social network can be represented as a graph structure, and the graph attention network can capture the characteristics of the interaction relationship between users and convert it into a social interaction feature vector. Assuming that the user's social interaction data is encoded as [0, 1, 1, 0], after being processed by the graph attention network, the resulting feature vector may be [0.6, 0.4].

[0120] Next, the system concatenates the user input feature vector, the device perception feature vector, and the social interaction feature vector to form a multimodal feature. For example, the concatenation of the above three feature vectors yields [0.2, 0.5, 0.1, 0.8, 0.1, 0.9, 0.6, 0.4]. Then, the system calculates the feature correlation of the multimodal feature through a multi-head attention mechanism. The multi-head attention mechanism can capture the correlation between different features. For example, the user input "play music" may be associated with the coffee shop in the current geographic location. Based on the feature correlation, the system generates an interaction pattern vector, such as [0.3, 0.7, 0.2].

[0121] In order to consider the impact of historical interaction data, the system uses a gated update mechanism to fuse the interaction pattern vector with the historical interaction sequence to obtain a fused interaction state. For example, the current interaction pattern vector [0.3, 0.7, 0.2] is fused with the historical interaction sequence [0.1, 0.6, 0.3] to obtain a fused interaction state [0.25, 0.65, 0.25].

[0122] Finally, the system builds a two-branch policy network based on the fused interaction state. The first branch is used to generate personalized recommendation strategies, such as recommending music that the user may like or nearby coffee shops. The second branch is used to generate real-time interaction solutions, such as providing navigation services or recommending nearby activities based on the user's real-time location and input information. The personalized recommendation strategy and the real-time interaction solution are combined to obtain an intelligent interaction strategy. For example, the strategies of recommending music and recommending nearby coffee shops are combined to form the final intelligent interaction strategy. The intelligent interaction strategy is then sent to the corresponding edge computing node.

[0123] Through the fusion of multimodal data and the generation of intelligent strategies, the system of the present invention can more accurately understand user intentions, provide more personalized and timely services, and thus significantly improve the user interaction experience; the system can predict the user's next operation based on the user's historical interaction data and real-time context, and provide corresponding suggestions and solutions, thereby simplifying the user operation process and improving interaction efficiency; the system can establish a user portrait based on the user's multimodal interaction data, and provide personalized recommendations and services based on the user portrait to meet the user's personalized needs.

[0124] In an optional embodiment,

[0125] A dual-branch strategy network is constructed based on the fused interaction state, wherein the first branch of the dual-branch strategy network is used to generate a personalized recommendation strategy, and the second branch of the dual-branch strategy network is used to generate a real-time interaction plan. The steps of strategically combining the personalized recommendation strategy and the real-time interaction plan to obtain an intelligent interaction strategy include:

[0126] A shared coding layer is constructed by using a multi-layer perceptron, and the fused interaction state is input into the shared coding layer to obtain coding features, and the coding features are used for two parallel branches of the dual-branch strategy network;

[0127] The first branch of the dual-branch strategy network uses an Actor-Critic architecture to generate a personalized recommendation strategy; wherein the Actor network includes a multi-layer strategy neural network, generates an Actor hidden state based on the encoding feature, and outputs a recommended action probability distribution as a preliminary personalized recommendation strategy through a softmax activation function; the Critic network includes a multi-layer value neural network, generates a Critic hidden state based on the encoding feature and outputs a state value estimate, calculates a temporal difference error using a temporal difference learning method, updates the Actor network parameters and the Critic network parameters based on the temporal difference error, introduces an entropy regularization term to construct an Actor network loss function to optimize the preliminary personalized recommendation strategy to obtain a final personalized recommendation strategy;

[0128] The second branch of the dual-branch strategy network uses a conditional variational autoencoder to generate a real-time interaction plan, including inputting the encoding feature into an encoder to obtain a mean vector and a variance vector as latent variable distribution parameters, sampling latent variables based on the latent variable distribution parameters through a reparameterization method, splicing the latent variables with the encoding feature, and then sequentially performing time series modeling through a decoding multi-layer perceptron and a gated recurrent unit, and inputting the hidden state of the gated recurrent unit into a generator network to obtain an interaction instruction to form a real-time interaction plan;

[0129] Construct a strategy combination module, the strategy combination module includes: a mutual information calculation unit, used to calculate the mutual information between the personalized recommendation strategy and the real-time interaction plan; a joint optimization unit, used to construct a joint optimization target using the mutual information as a regularization term; a strategy fusion unit, used to calculate the interaction strategy weight coefficient based on the joint optimization target, perform weighted combination of the personalized recommendation strategy and the real-time interaction plan, and generate an intelligent interaction strategy.

[0130] Exemplarily, first, prepare the interaction data. Collect user historical behavior data, such as clicks, browsing, purchases, etc., and context information, such as time, location, device, etc. Preprocess these data, such as cleaning, conversion, normalization, etc., to form a fused interaction state. For example, user Xiao Ming browsed product A and product B using a mobile phone at 2 pm on Monday, and purchased product A. The fused interaction state can be expressed as [Monday, 2 pm, mobile phone, product A browsed, product B browsed, product A purchased].

[0131] Next, build a shared encoding layer. Use a multi-layer perceptron to encode the fused interaction state and extract feature representation. For example, you can build a fully connected neural network with two hidden layers, each containing 64 neurons, and use the ReLU activation function. Input the fused interaction state into the network to obtain a 64-dimensional encoding feature.

[0132] Then, the first branch of the two-branch policy network is constructed to generate personalized recommendation strategies. This branch adopts the Actor-Critic architecture. The Actor network is a multi-layer policy neural network that inputs encoded features and outputs the probability distribution of recommended actions. For example, a fully connected network with two hidden layers, each containing 64 neurons, can be constructed, the number of neurons in the output layer is equal to the number of recommended actions, and the probability distribution is output using the softmax activation function. The Critic network is a multi-layer value neural network that inputs encoded features and outputs state value estimates. For example, a network with a similar structure to the Actor network can be constructed, with an output layer containing one neuron and outputting a scalar value of the state value. The temporal difference learning method is used to calculate the temporal difference error, and the parameters of the Actor network and the Critic network are updated based on the error. An entropy regularization term is introduced in the loss function of the Actor network to encourage the exploration of more possible recommendation strategies. For example, after Xiao Ming’s encoded features are input into the Actor network, the output is [Product C: 0.5, Product D: 0.3, Product E: 0.2], which means that the probabilities of recommending products C, D, and E are 0.5, 0.3, and 0.2, respectively.

[0133] At the same time, the second branch of the two-branch policy network is constructed to generate real-time interactive solutions. This branch uses a conditional variational autoencoder. The encoded features are input into the encoder to obtain the mean vector and variance vector as the parameters of the latent variable distribution. The latent variables are sampled from the distribution using the reparameterization method. After the latent variables and the encoded features are spliced, they are input into the decoder. The decoder consists of a multi-layer perceptron and a gated recurrent unit for time series modeling. The hidden state of the gated recurrent unit is input into the generator network to generate interactive instructions to form a real-time interactive solution. For example, after the encoded features are input into the encoder, the mean vector [0.2, 0.5] and the variance vector [0.1, 0.2] are obtained. The latent variable [0.3, 0.6] is sampled from it, spliced ​​with the encoded features, and input into the decoder. The decoder outputs a series of interactive instructions, such as ["show pictures of product C", "play videos of product C", "provide coupons for product C"], to form a real-time interactive solution.

[0134] Finally, a strategy combination module is constructed. This module contains three units: mutual information calculation unit, joint optimization unit and strategy fusion unit. The mutual information calculation unit calculates the mutual information between the personalized recommendation strategy and the real-time interaction scheme. The joint optimization unit adds the mutual information as a regular term to the joint optimization target. The strategy fusion unit calculates the interaction strategy weight coefficient based on the joint optimization target, performs a weighted combination of the personalized recommendation strategy and the real-time interaction scheme, and generates an intelligent interaction strategy. For example, the calculated mutual information between the personalized recommendation strategy and the real-time interaction scheme is 0.8. This value is added to the joint optimization target, and the weight coefficient is optimized to [0.7, 0.3]. The final intelligent interaction strategy is: recommend product C with a weight of 0.7, and execute interactive schemes such as ["show pictures of product C", "play videos of product C", "provide coupons for product C"] with a weight of 0.3.

[0135] The present invention meets user needs more accurately and provides a smarter interactive experience through personalized recommendations and real-time interactions; guides user behavior through interactive solutions to increase the probability of user purchases and other conversion behaviors; and allocates recommendation resources and interactive resources more effectively through intelligent strategies to improve overall efficiency.

[0136] Figure 2 Schematic diagram of the structure of a private domain live broadcast content distribution and visitor interaction system based on edge computing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0137] The first unit is used to receive a live content distribution request, generate a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy includes core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establish a distributed collaborative network based on the hierarchical deployment strategy, and send node deployment and resource configuration instructions to corresponding edge computing nodes;

[0138] The second unit is used for the regional layer nodes to collect the load status information of the edge computing nodes through the distributed collaborative network, adopt the federated learning method to fuse the load status information of the edge computing nodes, combine the LSTM time series prediction and attention mechanism to obtain the access popularity prediction value of the live content, and build an adaptive content distribution model; generate a content caching strategy based on the adaptive content distribution model, determine the distribution position of the content on the edge node according to the content caching strategy, and generate a corresponding content distribution routing solution;

[0139] The third unit is used for the access layer node to collect multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extract features from the multimodal interaction data, identify interaction patterns through an attention mechanism, and integrate historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions, and send the intelligent interaction strategies to corresponding edge computing nodes.

[0140] According to a third aspect of the embodiments of the present invention,

[0141] An electronic device is provided, comprising:

[0142] processor;

[0143] a memory for storing processor-executable instructions;

[0144] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0145] According to a fourth aspect of the embodiments of the present invention,

[0146] 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.

[0147] 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.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for private domain live broadcast content distribution and visitor interaction based on edge computing, characterized in that: include: Receive a live content distribution request, generate a hierarchical deployment strategy for edge computing nodes, the hierarchical deployment strategy includes core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; establish a distributed collaborative network based on the hierarchical deployment strategy, and send node deployment and resource configuration instructions to corresponding edge computing nodes; The regional layer nodes collect the load status information of the edge computing nodes through the distributed collaborative network, fuse the load status information of the edge computing nodes using the federated learning method, and obtain the access popularity prediction value of the live content by combining the LSTM time series prediction and the attention mechanism to build an adaptive content distribution model; generate a content caching strategy based on the adaptive content distribution model, determine the distribution location of the content on the edge node according to the content caching strategy, and generate a corresponding content distribution routing solution; The access layer node collects multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extracts features from the multimodal interaction data, identifies interaction patterns through an attention mechanism, and integrates historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions, and sends the intelligent interaction strategies to corresponding edge computing nodes; Receive a live content distribution request, and generate a hierarchical deployment strategy for edge computing nodes, wherein the hierarchical deployment strategy includes core layer nodes for global resource scheduling, regional layer nodes for content distribution optimization, and access layer nodes for user interaction; The steps of establishing a distributed collaborative network based on the hierarchical deployment strategy and sending node deployment and resource configuration instructions to corresponding edge computing nodes include: Receive a live content distribution request, the live content distribution request including an anchor identifier, a live broadcast type, and information about an expected audience size; generate a hierarchical deployment strategy using a hierarchical clustering algorithm according to a geographic location, a network topology, and computing resources; Based on the hierarchical deployment strategy, a two-phase commit protocol is adopted to establish a distributed collaborative network, including: sending a pre-deployment message containing node deployment and resource configuration to all layer nodes in the preparation phase, and issuing a deployment instruction after receiving confirmation information from all layer nodes in the commit phase; The functions of nodes at each layer are configured based on the deployment instructions, including: the core layer nodes establish a constraint optimization model for resource allocation based on a dynamic programming algorithm, and send the optimized resource allocation results to the regional layer nodes; the regional layer nodes use a minimum spanning tree algorithm based on the optimized resource allocation results to build an edge content distribution tree, and the edge weights of the edge content distribution tree are obtained by weighted calculation of network bandwidth, delay and jitter; the access layer nodes implement load balancing distribution of user requests based on a consistent hashing algorithm and a virtual node mechanism.

2. The method according to claim 1, characterized in that The core layer nodes establish a constrained optimization model for resource allocation based on a dynamic programming algorithm, and send the optimized resource allocation results to the regional layer nodes; the regional layer nodes use a minimum spanning tree algorithm to construct an edge content distribution tree based on the optimized resource allocation results, and the edge weights of the edge content distribution tree are calculated by weighting network bandwidth, delay, and jitter; The steps of implementing load balancing distribution of user requests based on the consistent hashing algorithm and the virtual node mechanism at the access layer node include: A resource allocation constraint optimization model for core layer nodes is constructed based on a dynamic programming algorithm and a Lagrange multiplier method. The resource allocation constraint optimization model includes allocation parameters for computing resources, storage resources, and bandwidth resources. An optimized resource allocation plan is obtained by solving an extreme value equation under resource allocation constraint conditions. The optimized resource allocation plan is sent to regional layer nodes. The regional layer node constructs an edge content distribution tree based on the optimized resource allocation scheme, and establishes an edge content distribution network topology using a minimum spanning tree algorithm. The edge weights of the edge content distribution tree are obtained by adaptive weighted calculation of bandwidth, delay and jitter values. The weight coefficients of the adaptive weighting are dynamically adjusted based on a distribution performance metric function using an online learning method, and the topological structure information of the edge content distribution tree is sent down to the access layer node; The access layer node uses a consistent hashing algorithm and a dynamic virtual node adjustment mechanism based on the topology information to achieve load balancing distribution of user requests. The dynamic virtual node adjustment mechanism adaptively adjusts the number of virtual nodes according to the actual load factor of the node, and adjusts the distribution ratio of user requests among the nodes by dynamically adjusting the virtual node distribution; A multi-level feedback adjustment mechanism is established, wherein the access layer nodes periodically collect load status information and feed it back to the regional layer nodes for adjusting the edge content distribution tree, and the regional layer nodes collect content distribution efficiency indicators and feed them back to the core layer nodes for optimizing resource allocation solutions.

3. The method according to claim 1, characterized in that The regional layer nodes collect the load status information of the edge computing nodes through the distributed collaborative network, fuse the load status information of the edge computing nodes using the federated learning method, and obtain the access popularity prediction value of the live content by combining the LSTM time series prediction and the attention mechanism. The steps of constructing the adaptive content distribution model include: The load status information includes computing load information, storage capacity information and network bandwidth information; A load feature extraction model is maintained locally at each edge computing node, the input of the load feature extraction model is the state sequence of the node within a preset time window, and the output is the load feature vector of the node; the load feature vector is input into the predictor to obtain a predicted load value, and a fusion loss function is constructed based on the mean square error between the predicted load value and the actual load and the cosine similarity between the current feature and the historical feature; each edge computing node independently trains a local load feature extraction model based on the fusion loss function, and regularly uploads the load feature extraction model parameters to the coordination node; the coordination node determines the weight according to the computing power and data quality of the edge computing node, and distributes the received load feature extraction model parameters to each edge computing node after weighted aggregation; The load feature vector is concatenated with the historical access record to construct an input sequence, and a bidirectional long short-term memory network is used to extract time series features; an attention mechanism is introduced to calculate the importance weights of different time steps, and the time series features are weighted and summed based on the importance weights to obtain a context vector; the context vector is processed through a fully connected layer to obtain an access heat prediction value; Based on the load feature vector and access popularity prediction value of the edge computing node, a multi-objective optimization method is used to weightedly combine the load balancing target, access delay target and bandwidth cost target to construct a comprehensive objective function; wherein the load balancing target is used to minimize the maximum value of the node load and capacity ratio, the access delay target is used to minimize the sum of delays for users to access the node, and the bandwidth cost target is used to minimize the sum of bandwidth costs for content transmission between nodes; and an adaptive content distribution model is obtained by solving the comprehensive objective function.

4. The method according to claim 1, characterized in that: The steps of generating a content caching strategy based on the adaptive content distribution model, determining the distribution location of the content at the edge node according to the content caching strategy, and generating a corresponding content distribution routing solution include: Generate a content caching strategy based on an adaptive content distribution model, wherein the content caching strategy is determined by calculating the matching degree between the content and the edge computing node, wherein the matching degree is calculated based on the load characteristic vector, processing capacity, and access popularity prediction value of the edge computing node, and when the matching degree exceeds a preset threshold, the corresponding content is allocated to the corresponding edge computing node for caching; Constructing a content node allocation matrix according to the content caching strategy; determining the distribution location of the content on the edge computing node based on the content node allocation matrix, executing a single allocation constraint on each content, and executing a capacity constraint on each edge computing node; A reachability graph is constructed based on the distribution position of the content in the edge computing nodes, wherein the reachability graph includes a set of edge computing nodes and connection relationships between the nodes; in response to a user content access request, a set of source nodes that cache the access request content and edge computing nodes that the user can directly access are determined from the content node allocation matrix, and a path cost from the edge computing node that the user can directly access to each node in the source node set is calculated, wherein the path cost is calculated based on the number of path hops and the average load rate of the nodes on the path; and a path with the minimum path cost is selected as a content distribution routing solution.

5. The method according to claim 1, characterized in that The access layer node collects multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extracts features from the multimodal interaction data, identifies interaction patterns through an attention mechanism, and integrates historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions. The steps of sending the intelligent interaction strategies to corresponding edge computing nodes include: Based on the bidirectional long short-term memory network, the user input data is subjected to temporal feature extraction to obtain a user input feature vector, based on the compressed convolutional network, the device perception data is subjected to spatial feature extraction to obtain a device perception feature vector, and based on the graph attention network, the social interaction data is subjected to interaction feature extraction to obtain a social interaction feature vector; The user input feature vector, the device perception feature vector and the social interaction feature vector are spliced ​​to form a multimodal feature, the feature correlation of the multimodal feature is calculated through a multi-head attention mechanism, and an interaction pattern vector is generated based on the feature correlation; the interaction pattern vector is fused with the historical interaction sequence by using a gated update mechanism to obtain a fused interaction state; A dual-branch strategy network is constructed based on the fused interaction state, wherein the first branch of the dual-branch strategy network is used to generate a personalized recommendation strategy, and the second branch of the dual-branch strategy network is used to generate a real-time interaction plan. The personalized recommendation strategy and the real-time interaction plan are strategically combined to obtain an intelligent interaction strategy.

6. The method according to claim 5, characterized in that A dual-branch strategy network is constructed based on the fused interaction state, wherein the first branch of the dual-branch strategy network is used to generate a personalized recommendation strategy, and the second branch of the dual-branch strategy network is used to generate a real-time interaction plan. The steps of strategically combining the personalized recommendation strategy and the real-time interaction plan to obtain an intelligent interaction strategy include: A shared coding layer is constructed by using a multi-layer perceptron, and the fused interaction state is input into the shared coding layer to obtain coding features, and the coding features are used for two parallel branches of the dual-branch strategy network; The first branch of the dual-branch strategy network uses an Actor-Critic architecture to generate a personalized recommendation strategy; wherein the Actor network includes a multi-layer strategy neural network, generates an Actor hidden state based on the encoding feature, and outputs a recommended action probability distribution as a preliminary personalized recommendation strategy through a softmax activation function; the Critic network includes a multi-layer value neural network, generates a Critic hidden state based on the encoding feature and outputs a state value estimate, calculates a temporal difference error using a temporal difference learning method, updates the Actor network parameters and the Critic network parameters based on the temporal difference error, introduces an entropy regularization term to construct an Actor network loss function to optimize the preliminary personalized recommendation strategy to obtain a final personalized recommendation strategy; The second branch of the dual-branch strategy network uses a conditional variational autoencoder to generate a real-time interaction plan, including inputting the encoding feature into an encoder to obtain a mean vector and a variance vector as latent variable distribution parameters, sampling latent variables based on the latent variable distribution parameters through a reparameterization method, splicing the latent variables with the encoding feature, and then sequentially performing time series modeling through a decoding multi-layer perceptron and a gated recurrent unit, and inputting the hidden state of the gated recurrent unit into a generator network to obtain an interaction instruction to form a real-time interaction plan; Construct a strategy combination module, the strategy combination module includes: a mutual information calculation unit, used to calculate the mutual information between the personalized recommendation strategy and the real-time interaction plan; a joint optimization unit, used to construct a joint optimization target using the mutual information as a regularization term; a strategy fusion unit, used to calculate the interaction strategy weight coefficient based on the joint optimization target, perform weighted combination of the personalized recommendation strategy and the real-time interaction plan, and generate an intelligent interaction strategy.

7. A private domain live broadcast content distribution and visitor interaction system based on edge computing, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to receive a live content distribution request and generate a hierarchical deployment strategy for edge computing nodes, wherein the hierarchical deployment strategy includes a core layer node for global resource scheduling, a regional layer node for content distribution optimization, and an access layer node for user interaction; Establish a distributed collaborative network based on the hierarchical deployment strategy, and send node deployment and resource configuration instructions to corresponding edge computing nodes; The second unit is used for the regional layer nodes to collect the load status information of the edge computing nodes through the distributed collaborative network, adopt the federated learning method to fuse the load status information of the edge computing nodes, combine the LSTM time series prediction and attention mechanism to obtain the access popularity prediction value of the live content, and build an adaptive content distribution model; generate a content caching strategy based on the adaptive content distribution model, determine the distribution position of the content on the edge node according to the content caching strategy, and generate a corresponding content distribution routing solution; The third unit is used for the access layer node to collect multimodal interaction data of users based on the distributed collaborative network, and the multimodal interaction data includes user input data, device perception data and social interaction data; extract features from the multimodal interaction data, identify interaction patterns through an attention mechanism, and integrate historical interaction data to generate intelligent interaction strategies, and the intelligent interaction strategies include personalized recommendation strategies and real-time interaction solutions, and send the intelligent interaction strategies to corresponding edge computing nodes.

8. 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 according to any one of claims 1 to 6.

9. 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 6 is implemented.

Images