Cloud edge collaborative caching method

Abstract

The invention relates to a cloud edge collaborative caching method, and belongs to the field of edge computing. According to the method, a core leader with the largest social influence is identified through a PageRank algorithm, a social relation network of the core leader is constructed, adjacent nodes are sent into a GCN to extract social association features, meanwhile, the dynamic preference of leader users is obtained through Transform Encoder, and a high-precision prediction sequence is generated through feature fusion. Aiming at the dynamic change of user preferences, a DDQN reinforcement learning algorithm with priority experience playback is adopted to carry out cache decision, and a reward function brought into social interaction feedback is set, so that a base station autonomously learns a cache updating strategy, and finally, the maximization of a long-term cache hit rate is realized, and the service request of a social group in a base station service area is met. According to the method, the cache can be intelligently and periodically updated, the backhaul link congestion can be relieved, and the user service quality is improved.

Description

Technical Field

[0001] This invention belongs to the field of edge computing and relates to a cloud-edge collaborative caching method. Background Technology

[0002] Edge caching is a core method for storage capabilities in edge computing. By caching data at edge nodes, it avoids data transmission back to the central cloud, allowing data distribution to be completed directly at the edge. This enables user requests to receive responses locally, significantly reducing transmission latency and backhaul load. If a node does not have the corresponding content, it will send a request to the remote cloud service center. However, due to the long distance transmission, this will cause video loading delays and stuttering, severely damaging the user's Quality of Experience (QoE).

[0003] Currently, multimedia traffic, represented by short videos, is experiencing explosive growth. This surge in traffic has placed unprecedented pressure on network infrastructure. When users frequently request popular content during peak network request periods, the network needs to repeatedly initiate requests, placing enormous repetitive transmission pressure on the core network. Therefore, to better design content caching to address this challenge, existing research has extensively studied user video content demand. Crane and Sornette et al. studied a large number of videos on YouTube to describe video request activity. Poisson processes can be used to accurately simulate most activities, while other activities can be categorized into three different patterns. Traverso et al. used access records from 60,000 YouTube end users to construct a realistic arrival mechanism for content requests. They then proposed a new model that allows users to capture the dynamics of content popularity locally and effectively analyze the caching system. Based on these theoretical findings, many researchers have attempted to construct algorithms and mechanisms for edge caching of video content. Tanzil et al. applied machine learning methods to construct an adaptive strategy for edge video content caching. The proposed strategy comprehensively considers user patterns and network attributes, aiming to select appropriate cache locations and cache sizes within the network.

[0004] With extremely limited storage resources at edge nodes, selecting and updating cached content to maximize cache hit rate and improve user QoE has become a pressing problem. Existing research mainly focuses on two directions: one is a global caching strategy based on content popularity, but it ignores the spatiotemporal locality of user preferences; the other is personalized caching based on machine learning, which predicts future requests by analyzing users' historical behavior. However, these methods generally face a core limitation: they treat users as isolated individuals, seriously neglecting the profound impact of complex social relationships between users on multimedia content consumption behavior. In short video platforms, users are not only consumers of content but also disseminators of content.

[0005] Therefore, a new caching method is urgently needed to solve the above problems. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a cloud-edge collaborative caching method that can intelligently and periodically update the cache to alleviate backhaul link congestion and improve user service quality.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A cloud-edge collaborative caching method is proposed, which establishes a proactive edge caching framework to pre-load potentially requested video content and intelligently update it periodically to improve user experience. The cloud center stores all requested video content and is responsible for responding to node requests. Users with complex social relationships at their physical node locations are considered as a group. Edge nodes store potentially requested content and deploy the caching computation framework. Edge nodes analyze the historical behavior data of community leaders and the influence characteristics of their friends using deep learning models to obtain initial node-predicted cached content. Reinforcement learning models the cache replacement problem as a Markov decision process, dynamically adjusting the strategy by maximizing long-term rewards.

[0008] The method specifically includes the following steps: S1: Grouping and Network Construction; S2: Finding social leaders; S3: Feature preprocessing; S4: Extraction of opinion leaders' own preference features; S5: Extraction of social relationship characteristics of community leaders; S6: Generate the initial prediction sequence; S7: Cache decision yields a high-precision predicted sequence.

[0009] Furthermore, step S1 specifically includes: treating all users within the coverage area of ​​a node as a social group; users under a base station have complex social relationships, such as friends, relatives, classmates, etc., and abstracting social relationships as edges and users as nodes, thereby constructing a social network graph under the edge nodes.

[0010] Furthermore, step S2 specifically includes: opinion leaders in social groups directly or indirectly influence the views and behaviors of ordinary individuals; the greater the in-degree of a user's node, the stronger the user's importance and social influence in the network; social influence is represented as:

[0011]

[0012] in, Indicates user u social influence For users u The base value of its own influence. For users v Activity coefficient For users v Its social influence; For users v The outgoing degree, i.e., the user v Number of followers; For users u The user group that is concerned This is the damping coefficient, with a default value of 0.85; the social relationships between nodes in a network are categorized as unidirectional or bidirectional based on the direction of the connection. Indicates the number of one-way edges. Indicates the number of bidirectional edges. As the normalization factor, For users u Activity coefficient.

[0013] Furthermore, step S3 specifically includes: sorting the user's historical video viewing records in chronological order, encoding different video types, converting the timestamp of the user's short video service request into a specific time, and normalizing the time of interaction with various services; the sequence of short videos viewed by the user is represented as follows:

[0014] in, This indicates the sequence of short videos a user has watched. Time period The video embedding representation is composed of various types of video components; The total number of types is 31. d For the embedded dimension; Time period Video types d The amount.

[0015] Building leader nodes in social networks Node features Due to type preference and average viewing ratio Composition, represented as: .

[0016] Furthermore, step S4 specifically includes: the user's requests at different times can reflect the user's dynamic preferences, the user's viewing sequence is represented as an embedding vector, and its location information is fed into the model to obtain the preference feature vector of the leader user.

[0017] Further, in step S4, the user's viewing sequence and location information are fed into the model to obtain the preference feature vector of the leader user. Specifically, this includes: feeding the user's viewing sequence and location information into the self-attention layer of the transformer. This layer uses a feedforward neural network to split the input vector into a query vector Q, a key vector K, and a value vector V. The self-attention mechanism multiplies each input Q by the K of other inputs (including itself) to finally obtain the matching V. Output of the head , ,..., The components are concatenated together, and a learnable linear projection matrix is ​​applied to the concatenated result. ,get ; Then, the output of multi-head attention The input to the module is added element-wise to alleviate gradient vanishing in deep networks. The result of the addition is then normalized layer by layer and fed into a feedforward neural network to finally obtain the preference representation of leading users. Represented as ,in , Representation layer normalization, Indicates regularization, Indicates the activation function; , This represents the weight matrix, which performs a linear transformation on the input features. , This is a bias term.

[0018] Furthermore, step S5 specifically includes: the social relationships between users will further influence the neighboring nodes of a user to take similar behaviors; a user's preference for services that may be requested in the future will be directly affected by the decisions of its neighboring nodes; information from each neighboring node is aggregated to generate a new feature representation; the embedded representation of the neighboring node users of the leader user in the network is calculated as follows:

[0019]

[0020] in, Indicates the first l Layer users u Embedded representation, i.e., information about adjacent nodes; Represents users in a social network u The set of adjacent nodes; For target users and adjacent nodes The strength of the connection, , All are learnable parameters. express Activation function It is the softmax activation function.

[0021] Furthermore, in step S5, information from each neighboring node is aggregated to generate a new feature representation. Specifically, this includes: using a lightweight GCN method to obtain the propagation of information from neighboring nodes; using an attention mechanism to characterize the influence of different friends; and weighting the influence of each neighboring node by its connection strength in the adjacency matrix. Through aggregation, the model learns the relationship structure between users and effectively integrates information from each neighboring node to generate a new feature representation.

[0022] Furthermore, step S6 specifically includes: fusing neighboring node information with the leader's dynamic preferences and the leader's close social relationships, and then generating initial pre-cached video content by passing the fused features through a multi-layer vector perceptron.

[0023] Furthermore, step S7 specifically includes: employing a reinforcement learning algorithm with priority experience replay for cache decision-making, setting a reward function with social interaction feedback, enabling the base station to autonomously learn cache update strategies, and ultimately maximizing the long-term cache hit rate; the reward function is defined as follows: ,in, For cache hit rate; Rewards for social interaction feedback are represented as follows:

[0024] in, This represents the number of times cached content within a time slice is actively shared by users via social links. This refers to the number of times cached content within a time slice is recommended to others by users with social connections. The total number of social links of leading users, For users n The strength of social relationships with leading users; To determine user satisfaction with cached content, the hit rate is... The value is 1 if it is 1, otherwise it is 0. Indicates base station cache content During the period t At the user's request, N For all users.

[0025] The beneficial effects of this invention are as follows: This invention utilizes a proactive edge caching framework based on cloud-edge collaboration. It pre-loads potentially requested video content and intelligently updates it periodically to improve user experience. The cloud center stores all requested video content and is responsible for responding to node requests. Users with complex social relationships at their physical node locations are considered a group. Edge nodes store potentially requested content and deploy the caching computation framework. Edge nodes analyze the historical behavior data of community leaders and the influence characteristics of their friends using deep learning models to obtain initial node-predicted cached content. Reinforcement learning models the cache replacement problem as a Markov decision process, dynamically adjusting the strategy by maximizing long-term rewards.

[0026] This invention identifies the most influential core leaders through the PageRank algorithm, constructs their social relationship networks, and feeds adjacent nodes into a Graph Convolutional Network (GCN) to extract social association features. Simultaneously, it utilizes a Transformer Encoder to obtain the dynamic preferences of these leaders, fusing features to generate high-precision prediction sequences. To address the dynamic changes in user preferences, a DDQN reinforcement learning algorithm with priority experience replay is employed for caching decisions. A reward function incorporating social interaction feedback is set, enabling the base station to autonomously learn cache update strategies, ultimately maximizing long-term cache hit rate and satisfying service requests from social groups within the base station's service area. This invention's intelligent and periodic cache updates can alleviate backhaul link congestion and improve user service quality.

[0027] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 A flowchart of a cloud-edge collaborative caching method provided in an embodiment of the present invention. Detailed Implementation

[0029] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] Please see Figure 1 This invention provides a cloud-edge collaborative caching method, specifically including the following steps: Step 1: Grouping and Network Construction (1) Users conduct activities within a certain range, and the geographical unit where they are located is regarded as a social attribute, dividing all users within the service range of a node into a social group.

[0031] (2) Users under the base station have complex social relationships. Each user is abstracted as a node, and the social relationships between users are abstracted as edges, thus forming a complex social network.

[0032] Step 2: Finding Social Leaders (1) Social relationships between nodes in a network are divided into one-way and two-way relationships according to the direction of the connection. Indicates the number of one-way edges. This indicates the number of bidirectional edges.

[0033] (2) Calculate the activity coefficient of each user and divide them into three levels: middle, high, and full, according to the number of times they watch videos. Set the damping coefficient to the standard value of 0.85.

[0034] (3) Calculation ,in, For users u The base value of its own influence. As the normalization factor, For users u Activity coefficient; (4) Calculation ,in, Indicates user u social influence For users v Activity coefficient For users v Its social influence; For users v The outgoing degree, i.e., the user vNumber of followers; For users u The user group that is concerned It is the damping coefficient, with a default value of 0.85.

[0035] Step 3: Feature Preprocessing (1) The user's request sequence consists of a limited number of video types that the user watches sequentially: To avoid assigning numerical meaning to labels in the model, the video type is encoded using LabelEncoder, converting string types into integers to achieve numerical representation of discrete types. Then, the discrete type encoding is converted into low-dimensional continuous vectors, and semantic relationships between types are learned through embedding layers. The total number of types is 31. d As an embedding dimension, the sequence of short videos watched by a user is represented as follows: , Time period Video types d The amount.

[0036] (2) Since users have different interaction durations with each item in the sequence, the interaction time (total viewing time / video duration) needs to be normalized using MinMaxScaler to scale it to the [0,1] interval, eliminating the influence of dimensions, and expressed as follows: ,in, Indicates the interaction time. This represents the longest view in the sequence. This represents the shortest video in the sequence.

[0037] (3) Calculation type preference Concatenate with the average viewing ratio vector, where type preference Based on the type of content viewed by users within a unit of time Frequency representation: ,in, Indicates user The total number of requests per unit time slice. This setting is used when a user has no request records within the time slice. The minimum value is 1e-8; For indicator functions, when the video type for Take 1 at the time; r This indicates the duration of the video.

[0038] (4) Calculate the average viewing ratio , User The mean of the normalized viewing percentages of all viewing records is expressed as: .

[0039] (5) Constructing leader nodes in social networks The node characteristics are composed of type preference and average viewing ratio. .

[0040] Step 4: Extraction of opinion leaders' own preference features (1) Divide user requests into several sequences per hour. Represent the content in the sequence as an embedding vector, which is obtained by combining video ID, content features, view rate, and timestamp, and adding position encoding to provide sequence order information to form the final input vector.

[0041] (2) The sequence and its position information are fed into the self-attention layer of the transformer. This layer uses a feedforward neural network to split the input vector into a query vector Q, a key vector K, and a value vector V. The self-attention mechanism multiplies each input Q by the K of other inputs (including itself) to finally obtain the matching V. Output of the head , ,..., The components are concatenated together, and a learnable linear projection matrix is ​​applied to the concatenated result. ,get .

[0042] (3) Output of multi-head attention The input to the module is added element-wise to alleviate gradient vanishing in deep networks. The result of the addition is then normalized layer by layer and fed into a feedforward neural network to finally obtain the preference representation of leading users. To support subsequent decisions on cached content, it is represented as ,in , Representation layer normalization, Indicates regularization, Indicates the activation function; , This represents the weight matrix, which performs a linear transformation on the input features. , This is a bias term.

[0043] Step 5: Extraction of social relationship characteristics of community leaders (1) To fully explore the preference representation of adjacent nodes of leader users, this part adopts a lightweight GCN method to obtain the propagation of information between adjacent nodes. Since the degree of intimacy is different, the influence of friends on leader nodes is also different. Therefore, an attention mechanism is used to characterize the influence of different friends. The influence of each adjacent node is weighted by its connection strength in the adjacency matrix. By aggregation, the model learns the relationship structure between users and effectively integrates the information from each adjacent node to generate new feature representations.

[0044] (2) The embedding representation of the neighbor set of the leader user in the computational network is as follows: ,in , Indicates the first l Layer users u Embedded representation, i.e., information about adjacent nodes; Represents users in a social network u The set of adjacent nodes, For target users and adjacent nodes The strength of the connection, , All are learnable parameters. express Activation function It is the softmax activation function.

[0045] Step 6: Generate the initial prediction sequence (1) Perform integration operations to generate fused user embedding representations, enabling dynamic user preference information. and neighboring node information Fusion during convolution, i.e. , ,in, Represents the normalization function; Indicates the first l Layer, User u Normalized fusion embedding; Indicates user u Friend neighbor embedding and dynamic preferences The fused feature representation, This represents the learnable weight matrix.

[0046] (2) The fused features are fed into the MLP, forward propagation is performed, a linear transformation is calculated using a fully connected layer, and then nonlinearity is introduced through ReLU activation. The linear transformation and activation operations are repeated to further extract higher-order features. The linear transformation is directly mapped to the predicted sequence length, and finally the predicted sequence of the set top-N length is output.

[0047] Step 7: Caching decision to obtain high-precision prediction sequence (1) Caching algorithms struggle to make optimal caching decisions. Therefore, based on the aforementioned social group preference prediction module, a dual-value network reinforcement learning algorithm (DDQN) is proposed to execute caching decisions, adapting to the dynamic and complex edge caching environment. In reinforcement learning, the base station acts as an agent, continuously interacting with the environment and learning how to better adapt. In time slots... At that time, the intelligent agent observes the state of the environment. And take action according to specific strategies. The agent will then receive a reward from the environment. The state will change from Transferred to The agent then takes new actions and receives new rewards. This learning process is repeated to maximize the agent's long-term discounted returns. Therefore, the node cache replacement problem is modeled as a Markov decision process, defining specific state spaces, action spaces, and reward functions.

[0048] (2) State Space: In the scenario of edge server cache replacement, the edge server needs to consider the impact of environmental factors when formulating caching strategies, and in different time slots Different strategies are formulated. The main factors influencing the caching strategy constitute the state space, including the current cached content sequence of the base station and the predicted sequence of social group preferences. ,in Indicates the base station caching strategy. This indicates the content of the group's prediction.

[0049] (3) Action space: Defines all possible actions that can be taken in a given state. For the base station buffer sequence decision problem, the agent's action is based on the time slot. The system state within the database provides the caching probability for each video, thus determining the video content that the base station needs to cache. The action is defined as... .

[0050] (4) Reward function: The agent rewards the system state. Make an action Appropriate rewards should be given afterward, and the training process should be adjusted in a direction that continuously increases the reward to maximize the expected reward. The social interaction feedback reward consists of two parts. The social interaction feedback reward quantifies the social interaction gain from three dimensions, reflecting the overall value of cached content to the social group under the base station: 1) social propagation gain, 2) group satisfaction balance, and 3) social relationship-weighted hit rate. Social propagation gain reflects the diffusion of content in the social network; the wider the propagation, the broader the potential demand coverage of the group, and caching this content can reduce duplicate requests from more edge nodes. Group satisfaction balance reflects the overall satisfaction of the cached content with the leader user's associated group. The social relationship-weighted hit rate assigns higher weight to the request hit rate of users with close social relationships to the leader user; the needs of these users are more likely to influence group preferences through social interaction, and their caching experience is more important for group stability. The reward function is defined as follows: ,in , This represents the number of times cached content within a time slice is actively shared by users via social links. This refers to the number of times cached content within a time slice is recommended to others by users with social connections. The total number of social links of leading users, For users n The strength of social relationships with leading users; To determine user satisfaction with cached content, the hit rate is... The value is 1 if it is 1, otherwise it is 0. Indicates base station cache content During the period t At the user's request, N For all users.

[0051] (5) DDQN maintains two independent sets of network parameters for the evaluation network and the target network. Both are fully connected neural networks, and the target network and the evaluation network have the exact same structure. (In time slots) At the same time, assess the network reception status. Output all actions The value has the largest The action with the highest value is selected as the best action. After the action is performed in the environment, the environment state transitions to... And provide instant rewards The state transition matrix generated by this process Save to replay experience pool To prevent the agent from repeatedly taking the same action in known states, thus failing to learn the optimal action for unknown states, linear decay is employed. The method involves environmental exploration. In a probability of... In such cases, the agent strictly takes the best action selected by the evaluation network.

[0052] (6) The core idea of ​​traditional experience replay is to uniformly and randomly sample quadruples from the experience pool for training. Although this method is simple, it may lead to some important experiences being ignored and unimportant experiences being sampled frequently, resulting in low learning efficiency. To promote model convergence, Prioritized Experience Replay (PER) is used as a mainstream off-policy deep reinforcement learning technique and integrated into the reinforcement learning algorithm. A batch of state transition matrices are sampled from the experience pool, and each transition matrix... In Into the evaluation network, and Input target network output corresponding The value is then used to obtain the loss function. The gradient descent method is used to update and evaluate the network parameters. The loss function is defined as follows: .

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A cloud-edge collaborative caching method, wherein the cloud center stores all requested video content and is responsible for responding to node requests; edge nodes store content that users may request and deploy a caching computing framework; characterized in that, Users with complex social relationships at a node's physical location are considered a social group; edge nodes use deep learning models to analyze the historical behavior data of social leaders under the node and the influence characteristics of the leader users' friends to obtain the initial node prediction cache content; Reinforcement learning models the cache replacement problem as a Markov decision process, dynamically adjusting the policy by maximizing long-term rewards; The method specifically includes the following steps: S1: Grouping and Network Construction; S2: Finding social leaders; S3: Feature preprocessing; S4: Extraction of opinion leaders' own preference features; S5: Extraction of social relationship characteristics of community leaders; S6: Generate the initial prediction sequence; S7: Cache decision yields a high-precision predicted sequence.

2. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S1 specifically includes: treating all users within the coverage area of ​​a node as a social group; users under a base station have complex social relationships, abstracting social relationships as edges and users as nodes, thereby constructing a social network graph under the edge nodes.

3. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S2 specifically includes: opinion leaders in social groups directly or indirectly influence the views and behaviors of ordinary individuals; the greater the in-degree of a user's node, the stronger the user's importance and social influence in the network; social influence is represented as: in, Indicates user u social influence For users u The base value of its own influence. For users v Activity coefficient For users v Its social influence; For users v The outgoing degree, i.e., the user v Number of followers; For users u The user group that is concerned It is the damping coefficient; the social relationships between nodes in a network are divided into unidirectional and bidirectional according to the direction of the connection. Indicates the number of one-way edges. Indicates the number of bidirectional edges. As the normalization factor, For users u Activity coefficient.

4. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S3 specifically includes: sorting the user's historical video viewing records in chronological order, encoding different video types, converting the timestamp of the user's short video service request into a specific time, and normalizing the time of interaction with various services; the sequence of short videos viewed by the user is represented as follows: in, This indicates the sequence of short videos a user has watched. Time period The video embedding representation is composed of various types of video components; Time period Video types d The amount; Building leader nodes in social networks Node features Due to type preference and average viewing ratio Composition, represented as: .

5. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S4 specifically includes: the user's requests at different times can reflect the user's dynamic preferences, the user's viewing sequence is represented as an embedding vector, and its location information is fed into the model to obtain the preference feature vector of the leader user.

6. The cloud-edge collaborative caching method according to claim 5, characterized in that, In step S4, the user's viewing sequence and location information are fed into the model to obtain the preference feature vector of the leader user. Specifically, this includes: feeding the user's viewing sequence and location information into the self-attention layer of the transformer. This layer uses a feedforward neural network to split the input vector into a query vector Q, a key vector K, and a value vector V. The self-attention mechanism multiplies each input Q by the other inputs K to finally obtain the matching V. Output of the head , ,..., The parts are concatenated together, and a learnable linear projection matrix is ​​applied to the concatenated result. ,get ; Then, the output of multi-head attention The input to the module is added element-wise to alleviate gradient vanishing in deep networks. The result of the addition is then normalized layer by layer and fed into a feedforward neural network to finally obtain the preference representation of leading users. Represented as ,in , Representation layer normalization, Indicates regularization, Indicates the activation function; , This represents the weight matrix, which performs a linear transformation on the input features. , This is a bias term.

7. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S5 specifically includes: the social relationships between users will further influence the neighboring nodes of a user to take similar behaviors; a user's preference for services that may be requested in the future will be directly affected by the decisions of its neighboring nodes; information from each neighboring node is aggregated to generate a new feature representation; the embedding representation of the neighboring node users of the leader user in the network is calculated as follows: in, Indicates the first l Layer users u Embedded representation, i.e., information about adjacent nodes; Indicates users in a social network u The set of adjacent nodes; For target users and adjacent nodes The strength of the connection, , All are learnable parameters. express Activation function It is the softmax activation function.

8. The cloud-edge collaborative caching method according to claim 7, characterized in that, In step S5, information from each neighboring node is aggregated to generate a new feature representation. Specifically, this includes: using a lightweight GCN method to obtain the propagation of information from neighboring nodes; using an attention mechanism to characterize the influence of different friends; weighting the influence of each neighboring node by its connection strength in the adjacency matrix; and using aggregation to enable the model to learn the relationship structure between users and integrate information from each neighboring node to generate a new feature representation.

9. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S6 specifically includes: fusing neighboring node information with the leader's dynamic preferences and the leader's close social relationships, and then generating initial pre-cached video content by passing the fused features through a multi-layer vector perceptron.

10. The cloud-edge collaborative caching method according to claim 1, characterized in that, Step S7 specifically includes: employing a reinforcement learning algorithm with priority experience replay for cache decision-making, setting a reward function with social interaction feedback, enabling the base station to autonomously learn cache update strategies, and ultimately maximizing the long-term cache hit rate; the reward function is defined as follows: ,in, For cache hit rate; Rewards for social interaction feedback are represented as follows: in, This represents the number of times cached content within a time slice is actively shared by users via social links. This refers to the number of times cached content within a time slice is recommended to others by users with social connections. The total number of social links of leading users, For users n The strength of social relationships with leading users; To determine user satisfaction with cached content, the hit rate is... The value is 1 if it is 1, otherwise it is 0. Indicates the base station cache content During the period t At the user's request, N For all users.

Images