Collaborative caching method based on gnn clustering in mobile edge network

Abstract

Social users in mobile edge network have highly similar content preferences and content requests, and common cooperative caching strategies do not place and update cache content of multiple base stations in view of the strong association of social users, resulting in a large amount of cache redundancy and higher content acquisition delay. Therefore, the application provides a cooperative caching method based on GNN clustering in a mobile edge network, first, user clustering is carried out on a social network by using E-GraphSAGE to establish interest communities with different content preferences; then, the cooperative caching problem for the interest community is converted into a mixed integer programming problem of maximizing the system MOS value to meet the QoE demand of the requesting user; finally, the MADDQN is used to solve the programming problem to obtain the optimal cooperative caching method. Simulation experiments show that, compared with the cooperative caching strategy based on K-Means clustering, the cooperative caching method has higher cache hit rate and lower content acquisition delay.

Description

Technical Field

[0001] This invention relates to computer network technology, specifically a collaborative caching method based on GNN clustering in mobile edge networks. Background Technology

[0002] The rapid development of mobile Internet technology and the widespread adoption of portable mobile terminals have resulted in global mobile data traffic exceeding 100 EB per month in 2023, putting enormous pressure on the backhaul bandwidth of mobile networks [1]. The traditional centralized mobile Internet architecture [2] migrates user content requests to the cloud center. The long-distance data transmission between the two leads to delays in the processing and distribution of content requests, making it difficult to meet the low-latency requirements for content acquisition and seriously affecting the user experience. Deploying computing and caching capabilities to distributed mobile edge networks (MEN) [3] is one of the key technologies to solve the above problems.

[0003] In MEN (Multi-User Environment), edge servers and base stations are configured with caching devices of varying capacities. The newly added caching function allows requested content to be pre-placed in the cache, enabling users to experience near-LAN response speeds, thus significantly reducing transmission latency and network congestion. Statistics show that over 80% of internet users repeatedly access 20% of information resources; content caching will effectively improve the quality of new data services such as web browsing, multimedia, and social networks. However, limited cache capacity and diverse content mean that base stations cannot promptly meet the content service needs of all associated users. Therefore, how to collaboratively cache popular content across multiple base stations to improve cache hit rates is a pressing issue that MEN needs to address when distributing content.

[0004] In recent years, a large number of online social networking applications have emerged, such as Facebook, Twitter, LinkedIn, WeChat, Tencent, and Sina Weibo. These social services have increasingly higher requirements for the QoS of content transmission speed and accuracy, which profoundly affects the design and implementation of collaborative caching strategies [4][5]. On the other hand, social users have highly similar content preferences. If collaborative caching is based on common interests, redundant data can be reduced and the cache hit rate of social groups can be enhanced. Therefore, researchers have explored more efficient collaborative caching strategies based on the clustering results of social users. Reference [6] proposed a socially aware edge caching strategy (SECS), which implements clustering by estimating the similarity between the interest domains of social users and realizes intra-cluster cache updates by content sharing performance. Compared with non-clustered collaborative caching, SECS has achieved better performance in terms of cache hit rate, cache cost, response latency, and control overhead. Reference [7] studied the optimal collaborative caching strategy for D2D communication in heterogeneous networks based on user mobility and social relationships, and theoretically derived a closed expression for the average system cost. After clustering social users using a Zipf distribution, the active caching of users within the group is optimized by minimizing the average system cost, thus reducing the content service cost for operators. The aforementioned literature, when establishing interest groups using traditional clustering methods, failed to fully consider the implicit non-Euclidean space characteristics of social networks. Mining multi-user relationships based on graph structures can help further improve clustering performance. Therefore, how to utilize graph structures to obtain more accurate user clustering to further improve collaborative caching performance remains an unsolved problem.

[0005] Graph Neural Network (GNN) [8] is an end-to-end deep learning model that uses graph structures to express the relationships between objects and mines features from graph data to complete various graph analysis tasks, such as node clustering [9], community detection and detection

[10] , user clustering

[11] , etc. In recent years, researchers have proposed many GNN methods to solve the problem of object clustering, which are mainly divided into two categories: clustering based on node information, such as the GraphSAGE algorithm

[12] which uses small-batch neighbor sampling and aggregation centered on all graph nodes to complete GNN training and obtain node clustering; and clustering based on edge information, such as the E-GraphSAGE algorithm

[13] which performs GraphSAGE on the two endpoints of each edge, connects the features of the two endpoints as edge embedding, and then completes GNN training through edge information transmission and aggregation to obtain edge clustering. Celebrity accounts on social networks such as Facebook and Sina Weibo have attracted the attention of tens of millions of fans. These fans have many similar interests and hobbies, which can provide a basis for user clustering. Reference

[14] designed a social network representation learning algorithm, which infers the user's status and role in a graph-structured social network by observing the attention of celebrity neighbors and using a gate mechanism. Experimental results on four datasets, Citeseer, Cora, Wisconsin and Cornell, show that the algorithm improves the clustering accuracy by 2% compared to the GraphSAGE-Mean algorithm. Reference

[15] proposed a social network node clustering algorithm based on graph coding networks. Each node integrates its own attributes and the interest correlation with neighboring nodes to extract high-level edge features, and uses this to achieve social user clustering. Compared with deep random walk models, logistic regression algorithms and graph convolutional network algorithms, the proposed algorithm improves the clustering accuracy on the Weibo dataset by 8%, 4.83% and 0.91%, respectively. The research work of the above references provides a theoretical and experimental basis for collaborative caching strategies for user clustering. Summary of the Invention

[0006] In mobile edge networks, social users have highly similar content preferences and requests. However, common collaborative caching strategies do not place and update cached content across multiple base stations based on the strong correlation among social users, resulting in significant cache redundancy and high content retrieval latency.

[0007] To address the aforementioned problems in existing technologies, this invention proposes a collaborative caching method based on GNN clustering to efficiently distribute content to social users.

[0008] The design steps of the method of this invention include:

[0009] First, E-GraphSAGE is used on social networks to cluster users and establish interest communities with different content preferences;

[0010] Next, the collaborative caching problem for interest communities is transformed into a mixed-integer programming problem that maximizes the system's MOS value to meet the QoE requirements of requesting users;

[0011] Finally, the optimal collaborative caching method is obtained by solving the planning problem using MADDQN.

[0012] Simulation experiments show that, compared with collaborative caching strategies based on K-Means clustering, the collaborative caching method proposed in this invention has a higher cache hit rate and lower content retrieval latency. Attached Figure Description

[0013] Figure 1 This is a scenario diagram of social user clustering in mobile edge networks.

[0014] Figure 2 It is a bipartite graph of a social user dataset.

[0015] Figure 3 This is a schematic diagram of the E-GraphSAGE model training process.

[0016] Figures 4(a) and 4(b) show the impact of the number of clusters on the performance of the collaborative cache, where: Figure 4(a) is the impact of the number of clusters on MOS; Figure 4(b) is the impact of the number of clusters on the cache hit rate.

[0017] Figures 5(a) and 5(b) show the impact of base station capacity on cooperative cache performance, where: Figure 5(a) shows the impact of base station capacity on MOS; Figure 5(b) shows the impact of base station capacity on cache hit rate.

[0018] Figures 6(a) and 6(b) show the impact of the number of base stations on the performance of the cooperative cache. Figure 6(a) shows the impact of the number of base stations on the MOS (Mean Offset Tolerance). Figure 6(b) shows the impact of the number of base stations on the cache hit rate.

[0019] Figures 7(a) and 7(b) show the impact of file type on collaborative caching performance in edge servers. Figure 7(a) shows the impact of file type on MOS, and Figure 7(b) shows the impact of file type on cache hit rate.

[0020] Figures 8(a) and 8(b) show the impact of the number of users on collaborative caching performance, where: Figure 8(a) shows the impact of the number of social users on MOS; Figure 8(b) shows the impact of the number of social users on cache hit rate. Detailed Implementation

[0021] 1 Overview

[0022] This invention proposes a collaborative caching strategy based on GNN clustering, which uses GNN clustering to form different interest communities among social users. The edge server is designed with this interest community-centric collaborative caching strategy. This social group-centric strategy can effectively reduce cache redundancy at each base station; simultaneously, it can enhance content distribution efficiency through multicast when spectrum resources are limited.

[0023] 2. Problem Description

[0024] For ease of explanation, this invention considers a cooperative caching strategy for an edge cell, such as... Figure 1 As shown, this MEN system consists of edge servers, base stations, and terminal social users. Social users are divided into different interest communities based on their social attributes, and different base stations can be associated with similar interest communities. For example, users in community A and community B, associated with two base stations, may prefer short videos and movies / TV shows, respectively. The edge server is responsible for caching the set of requested content from all users within its area. Base station set Responsible for collaborative caching of interest communities The system can receive requests from users and send them to other users in the community via cellular links using multicast. Neighboring base stations can also exchange information and share content.

[0025] against Figure 1 In the context of MEN user clustering, this invention studies the collaborative caching problem of popular content. Reference

[25] found after statistical analysis of a large number of user demands that the content access volume of users follows a Zipf distribution, that is, 80% of user requests come from 20% of cached content. However, the content access volume of users in MEN is huge, and even if 20% of the popular content is cached, it is still difficult to bear the limited cache space of the base station. Therefore, under the limited cache space, the first problem to be solved by this invention is to select the most valuable popular content for the same community users. On the other hand, although caching the most valuable popular content for each base station can improve the cache hit rate, it will lead to a large amount of cache redundancy and reduce the diversity of cached content. Therefore, how to coordinate multiple base stations to cache high-value popular content for the same community users is the second problem to be solved by this invention.

[0026] To address the two issues mentioned above, this invention designs a social user clustering strategy based on E-GraphSAGE and a collaborative caching strategy based on social user clustering, respectively, to improve the content service QoE of the MEN system.

[0027] 3. Social User Clustering Based on E-GraphSAGE

[0028] Figure 1Accurate social user clustering is the basis for multiple base stations to select and cache popular content. Therefore, this invention uses the E-GraphSAGE algorithm in graph neural networks (GNNs) to divide social users into different interest communities on a social user dataset.

[0029] 3.1 Preprocessing of Social User Data

[0030] Social user datasets contain rich graph structure information, mainly composed of ordinary user IDs, celebrity user IDs, and social attributes, including following, followers, gender, and interests. Utilizing the graph structure relationships of social networks in a non-Euclidean space, this invention transforms the social user clustering problem into an edge clustering problem. First, a batch of user information is constructed as follows... Figure 2 The two-part diagram shown ,in These represent the set of ordinary user nodes, the set of celebrity user nodes, and the edge set, respectively. Figure 2 Identify the source nodes as ordinary and celebrity user IDs respectively. , destination node Then, the feature vectors of social attributes are... The edge features correspond to the source and destination nodes. To facilitate analysis and computation, this invention transforms the bipartite graph of the social user dataset into a conventional graph. ,in Represents the set of vertices. Represents the set of edges.

[0031] 3.2 E-GraphSAGE Algorithm

[0032] As a classic graph neural network, E-GraphSAGE captures graphs... The topological structure and edge features are used to implement edge embedding and edge clustering. For example... Figure 3 As shown, the training process of the E-GraphSAGE algorithm can be summarized into the following three steps.

[0033] Step 1: Random Sampling. First, sample the neighborhood of any node v in the node set V. Defined as a set of fixed size and uniformly sampled (See Figure 3 (Step 1 diagram). Then, uniformly select 8 neighboring nodes with 2 hops from the sampling set for edge batching. The kth iteration

[0034] (1)

[0035] Step 2: Edge Feature Aggregation. Initialize node feature vectors. Its dimension is equal to the dimension of the edge eigenvectors. Figure 3 In the second step diagram, the first... Layers, using aggregation functions The sampled neighborhood feature vector of the (k-1)th layer Aggregate to node v, i.e.

[0036] (2)

[0037] Step 3: Embedding edge information. Aggregating information. Embedded information with the (k-1)th layer nodes Multiply by the weight matrix after splicing Then, according to equation (3), the activation function Update node embedding

[0038] (3)

[0039] Among the symbols This represents the concatenation. After K aggregation iterations, the final edge embedding representation is... It is the concatenation of the embedding vectors of two nodes, that is...

[0040] (4)

[0041] It is noted that the original edge features may suffer loss during average aggregation and cannot be well represented in the final edge embedding. To address this issue, this invention concatenates the two endpoint embeddings of edge uv and the original edge features according to equation (5).

[0042] (5)

[0043] at last, The class probabilities are obtained after the Softmax layer, and then social user clustering is completed, such as... Figure 3 The diagram for step three is shown below.

[0044] Through E-GraphSAGE's social user clustering, content requesting users within edge communities can be grouped into categories such as... Figure 1 The different interest communities shown lay the foundation for collaborative caching of content services for social users.

[0045] 4. MEN System Model Based on Social User Clustering

[0046] Common collaborative caching strategies calculate content retrieval latency centered on individual users, aiming to minimize the total latency across multiple users. However, observing the long-term request records of resident users within a base station's coverage area reveals that these users have relatively stable interests and highly similar social attributes. Providing precise group services to users in similar interest communities can effectively reduce cache redundancy and communication costs. Section 3 uses E-GraphSAGE to divide social users into different interest communities; the following section will further analyze the MEN system model based on social user clustering.

[0047] 4.1 Demand Model

[0048] The internet currently contains a massive amount of content, some of which is frequently accessed by a large number of users, while other content is rarely accessed. Content popularity reflects users' liking for the content, and caching the most popular content on the internet will significantly improve users' content hit rate. Typically, content... The popularity follows a Zipf distribution.

[0049] (6)

[0050] in Content Rank in a descending popularity sequence; Indicates average popularity; This represents the offset index of the Zipf distribution. The larger the value, the higher the user demand.

[0051] As mentioned earlier, users in similar interest communities show a high degree of similarity in their liking for certain popular content. Figure 1 In the context of interest communities The collection of popular content favored by users is Then, based on the content accessed by the user Average request rate and average access duration can be defined as the following interest preferences.

[0052] (7)

[0053] in , These represent the views of all users within the interest community c on the content. The number of requests and the duration of access; , These represent the total number of requests and the total access duration from all users within interest community c to all cached content on the edge server, respectively. and All of these represent weighting factors.

[0054] Equations (6) and (7) show that content popularity and interest preferences influence interest communities. Internal users' content To address the demand, we might define a value function to measure the magnitude of content demand.

[0055] (8)

[0056] in and All represent weighting factors. Clearly, the higher the value of the content, the higher the probability it will be requested by users. Therefore, we define an interest community. Internal user content set The average value function is

[0057] (9)

[0058] 4.2 Average Transmission Delay Model

[0059] If users in similar interest communities request the same content within a time slot, the base station can significantly reduce communication resource consumption by sending the requested content via multicast. Therefore, this invention adopts multicast for downlink data transmission. Interest Communities Multiple users within the area send signals to the local base station. After initiating a request for the same content, if the local base station has cached the requested content, it will send it directly to the user; otherwise, if multiple neighboring base stations have cached the requested content, the local base station will obtain the content from the nearest neighboring base station and forward it to the user; otherwise, it will obtain the content from the edge server and forward it to the user. Let's define... , Each represents the content Cached and uncached data at the local base station; Each represents the content Cached or uncached at neighboring base stations.

[0060] Assuming base station To community users Base stations to base station Edge server to base station The Euclidean distance and the data transmission speed per unit are respectively and The corresponding transmission delays are respectively

[0061] (10)

[0062] (11)

[0063] (12)

[0064] in Indicates the size of the content.

[0065] Considering that base stations use multicast for content distribution, therefore base stations Interest communities within the coverage area Content accessed by all users The transmission delay can be expressed as

[0066] (13)

[0067] in Indicates base station To users Distribute content The transmission delay. Similarly, the base station Content is obtained from the nearest neighboring base station or edge server. Then forwarded to interest-based communities The transmission delays for internal users are respectively Thus, interest communities Requesting content from users The transmission delay is

[0068] (14)

[0069] Figure 1 Interest communities exist within the coverage areas of multiple base stations in China. This will combine the different content requested by all users into a set. Then the average transmission latency for these users to obtain content is

[0070] (15)

[0071] in Represents a collection of requested content Number of contents in the middle Indicates related interest communities The number of base stations.

[0072] 5 Collaborative Caching Strategy Based on Social User Clustering

[0073] 5.1 Modeling the Collaborative Caching Problem

[0074] Based on the above demand model and average transmission latency model, this invention designs a collaborative caching strategy based on social user clustering in the MEN scenario. Inspired by the widely used QoE metric, the Mean Opinion Score (MOS) model

[26] can be used to measure the quality of content services. As can be seen from equation (15), interest communities Chinese user requests The MOS score for all content in the content is

[0075] (16)

[0076] Where the constant , Make From equation (16), we can see that... The higher the score, the more interest-based communities there are. The higher the QoE of internal users.

[0077] Assumption Figure 1 The collection of content requests from all interested community users is as follows And the cached content set of all base stations is The objective of this invention is to maximize the system's MOS value to meet the QoE requirements of requesting users in the edge region. Therefore, the collaborative caching problem is transformed into the following optimization problem.

[0078] (17)

[0079] in Interest-based communities The minimum transmission latency required by the user. Indicates that the cache is stored at the base station. A collection of content, Indicates base station The currently available cache capacity; constraint C1 indicates that the transmission delay of the content distributed by the base station is not greater than the minimum transmission delay requirement; constraint C2 ensures that the total amount of cached content of the base station is not greater than its rated cache capacity; constraints C3 and C4 represent the 0-1 decision of content caching;

[0080] The mixed integer programming problem (17) can be transformed into a Markov Decision Process (MDP). This invention utilizes Multi-Agent Double Deep Q-Network (MADDQN)

[16] to accurately solve the MDP problem in order to obtain the optimal collaborative caching strategy.

[0081] 5.2 Solving the Collaborative Caching Problem

[0082] The edge server uses MADDQN to interact with the environment through trial and error, seeking an optimal strategy by maximizing the cumulative reward. Typically, the MADDQN-based collaborative caching problem can be represented as a triple. ,in Representing the state space, Represents the action space. These represent the immediate reward value gained from completing an action, and their specific descriptions are as follows.

[0083] State-space agents Before selecting an action, you need to understand the base station. Status information and interest communities under this base station The content requested by the Chinese user Therefore, the state space medium state It can be represented as

[0084] (18)

[0085] in .

[0086] Action space agent m received interest community Chinese user request content Afterwards, if the local base station Or there is content in the cache of adjacent base stations. Then the action This indicates that no cache placement or update will be performed; conversely, the action indicates that no cache placement or update will be performed. Indicates that the base station Replace a certain content in the cache with the content. .

[0087] Reward function environment state Below, intelligent agents Execute action Then proceed to the next state And receive instant rewards. The optimization objective of this paper is to maximize the user's QoE, therefore the MOS value is set to the reward value.

[0088] (19)

[0089] Where the constant Make .

[0090] state space To the action space The mapping constitutes a strategy : .

[0091] The MADDQN model contains multiple agents, among which... For base stations Determine a collaborative caching strategy for request content across all interest communities within the coverage area. Define the current state vector. intelligent agent Use strategy Take action The action-state value function is

[0092] (20)

[0093] in Indicates base station A collection of all interest communities within the coverage area. This is the discount factor. According to the Bellman Equation, The update formula for the function is:

[0094] (twenty one)

[0095] in It is the learning rate that controls the learning speed.

[0096] This invention uses MADDQN to find the optimal collaborative caching strategy.

[0097] That is, intelligent agent Update the main network parameters To approximate the optimal strategy The corresponding Q value

[0098] (twenty two)

[0099] To maintain the independence of the sampled data, MADDQN establishes an experience replay mechanism, which replays the experience gained by agent m. Experiences are stored in an experience replay pool to record each interaction with the environment; when the pool is full, new experiences randomly replace older ones. On the other hand, to avoid overestimating the Q-value during training, MADDQN constructs a main network for selecting actions and a target network for estimating the value function, assuming that the state-action value functions of the two networks are respectively... and ,in This represents the target network parameter.

[0100] The training process of the MADDQN model can be summarized as follows: First, the agent m randomly samples mini-batch samples from the experience replay pool and inputs them into its main network and target network, and calculates the Q-values ​​of the two networks using forward propagation. Next, a loss function is established.

[0101] (twenty three)

[0102] Backpropagation is performed on the main network to update the main network parameters. Specifically, the gradient of the loss function in equation (23) is calculated.

[0103] (twenty four)

[0104] Get parameters Update formula

[0105] (25)

[0106] in This represents the learning rate. Finally, every Z steps, agent m copies the parameters of the main network to the target network, i.e. Used to update the parameters of the target network. Based on the above discussion, a collaborative caching strategy based on social user clustering is presented below.

[0107] Algorithm 1: Collaborative Caching Strategy Based on Social User Clustering

[0108]

[0109] 6. Simulation Experiments and Result Analysis

[0110] 6.1 Simulation Environment Setup

[0111] This invention verifies on the Python platform that the proposed GNN-MADDQN can maximize the QoE of user content services. Assume that one edge server is deployed in the center of a 3km×3km area, and the number of base stations evenly distributed around it ranges from 3 to 10. The edge server caches 10 types of files; the maximum cache capacity of each base station ranges from 4 to 10GB. The number of users extracted from the dataset

[17] ranges from 20 to 140, and the number of interest communities formed by clustering ranges from 2 to 40. Then, each interest community user is randomly assigned to multiple base stations. The content requests of users in the interest community are randomly generated. Other simulation parameters are shown in Table 1.

[0112] Table 1 Parameter Settings

[0113]

[0114] To illustrate the performance of GNN-MADDQN collaborative caching, this invention sets up the following three benchmark algorithms:

[0115] (1) Random Algorithm: No social users are clustered, and the agent randomly caches the content of each base station.

[0116] (2) GNN-DDQN: GNN is used to cluster social users, and the agent uses DDQN to obtain the collaborative caching strategy.

[0117] (3) K-Means-MADDQN: K-Means is used to cluster social users, and the agent uses MADDQN to obtain the collaborative caching strategy.

[0118] 6.1 Simulation Result Analysis

[0119] Figures 4(a) and 4(b) illustrate the impact of the number of clusters on collaborative caching performance. As the number of clusters increases, the MOS value and cache hit rate of all four algorithms show a trend of first increasing and then decreasing. Random, GNN-DDQN, and GNN-MADDQN achieve their highest MOS value and cache hit rate with a cluster size of 10, while K-Means-MADDQN achieves its highest with a cluster size of 16. This demonstrates that the system cannot provide optimal content service quality when the number of clusters is too high or too low. To achieve higher caching performance, the following experiments all use a cluster size of 10.

[0120] Figures 5(a) and 5(b) illustrate the impact of base station capacity on collaborative caching performance. The MOS value and cache hit rate of all four algorithms increase continuously as the base station capacity increases from 4GB to 10GB, with GNN-MADDQN showing the most stable increase. This indicates that even when base station memory resources are occupied by other applications and the available memory resources for caching are limited, each agent in GNN-MADDQN can set the best collaborative caching strategy for the corresponding base station. Random does not consider the correlation characteristics between social users and caches base station content arbitrarily; therefore, it can only achieve a good content service QoE when base station memory resources are abundant and a large amount of popular content is cached. GNN-DDQN and K-Means-MADDQN consider social user clustering, thus their collaborative caching performance is superior to Random.

[0121] Figures 6(a) and 6(b) illustrate the impact of the number of base stations on collaborative caching performance. When the number of base stations increases from 3 to 10, the number of neighboring base stations for each base station increases. At this point, multiple neighboring base stations can collaboratively cache a wider variety of popular content, effectively increasing the total amount of content in the system. When multiple users request different types of content, the local base station is highly likely to obtain the requested content from itself and its neighbors, thereby improving the MOS value and cache hit rate. GNN can achieve better user clustering results than K-Means, so the MOS values ​​and cache hit rates of GNN-DDQN and GNN-MADDQN are higher than those of K-Means-MADDQN. Although K-Means-MADDQN uses MADDQN to solve the collaborative caching strategy, its MOS value is on average 5.4% and 9.0% lower than that of GNN-DDQN and GNN-MADDQN due to incomplete clustering accuracy. This indicates that the clustering results of social users significantly affect collaborative caching performance.

[0122] Figures 7(a) and 7(b) illustrate the impact of file types on collaborative caching performance in edge servers. The more file types an edge server has, the more popular content the base station needs to collaboratively cache, leading to higher caching pressure. While GNN-DDQN's collaborative caching strategy using DDQN output is slightly worse than MADDQN's, its MOS value is lower for different file types, indicating that content retrieval latency needs further optimization. GNN-MADDQN's social user clustering and caching strategies are superior to the other three algorithms, thus achieving optimal content service quality even when faced with changes in popular content.

[0123] Figures 8(a) and 8(b) illustrate the impact of the number of social users on collaborative caching performance. When the number of clusters is fixed, different numbers of social users will result in different clustering outcomes, thus affecting subsequent collaborative caching strategies. As the number of social users increases from 40 to 140, the types and quantities of requested content in the system increase accordingly, posing a challenge to base station caching. In Figures 8(a) and 8(b), the MOS values ​​and cache hit rates of the four algorithms all decrease, but the decreases for GNN-MADDQN and GNN-DDQN are smaller, while the decreases for Random and K-Means-MADDQN are both greater than 14.5%. Combined with the experimental results in Figures 6(a) and 6(b), this further demonstrates that better GNN clustering results allow a larger number of users with the same interests to obtain requested content faster using multicast technology, thereby alleviating the caching pressure on popular content caused by the increase in the number of users.

[0124] 7. Conclusion

[0125] To improve the efficiency of content retrieval for social users in a MEN (Multi-User Environment), this paper proposes a collaborative caching strategy based on GNN clustering. Leveraging the strong correlations among social users, this strategy uses GNN for precise clustering to identify different types of interest communities, and then uses MADDQN to solve for the collaborative caching strategy specific to these interest communities. Compared to traditional collaborative caching strategies that consider individual users, designing a collaborative caching strategy oriented towards social groups with interest communities as the basic unit effectively reduces system cache redundancy and improves content distribution efficiency. Simulation results demonstrate that, under different network environments, GNN clustering can assist MADDQN in achieving superior collaborative caching performance compared to Random, GNN-DDQN, and K-Means-MADDQN.

[0126] References

[0127] [1]Nishida R, Shimokawa M, Sanada K, et al. A station grouping method considering heterogeneous traffic and multiple data rates for IEEE 802.11 ahnetworks with non-uniform station deployment[C] / / 2022 IEEE 95th VehicularTechnology Conference: (VTC2022- Spring). IEEE, 2022: 1-5.

[0128] [2]Phuttharak J, Loke S W. A review of mobile crowdsourcingarchitectures and challenges: Toward crowd-empowered Internet-of-Things[J]. IEEE Access, 2018, 7: 304-324.

[0129] [3]Nomikos N, Zoupanos S, Charalambous T, et al. A survey onreinforcement learning-aided caching in heterogeneous mobile edge networks[J]. IEEE Access, 2022, 10: 4380-4413.

[0130] [4]Jiang L, Shi L, Liu L, et al. An efficient evolutionary userinterest community discovery model in dynamic social networks for internet of people[J]. IEEE Internet of Things Journal, 2019, 6(6): 9226-9236.

[0131] [5]Chung J M. Content Delivery Network (CDN) Technology[M] / / EmergingMetaverse XR and Video Multimedia Technologies: Modern Streaming andMultimedia Systems and Applications. Berkeley, CA: Apress, 2022: 251-277.

[0132] [6] Jia Shu, Zhou Zhi, Li Wenlin, et al. Socially-aware edge caching strategy for video resources in 5G ultra-dense networks [J]. Mobile Information Systems, 2021, 2021:1-14.

[0133] [7]Shan G, Zhu Q. Sociality and mobility-based caching strategy for device-to-device communications underlying heterogeneous networks[J]. IEEEAccess, 2019, 7: 53777-53791.

[0134] [8]Zhou J, Cui G, Hu S, et al. Graph neural networks: A review of methods and applications[J]. AI open, 2020, 1: 57-81.

[0135] [9]Li C, Chen H, Li T, et al. A stable community detection approachfor complex network based on density peak clustering and label propagation[J]. Applied Intelligence, 2022, 52(2): 1188-1208.

[0136]

[10] Srinivasavaradhan S R, Nikolopoulos P, Fragouli C, et al. Dynamicgroup testing to control and monitor disease progression in a population[C] / / 2022 IEEE International Symposium on Information Theory (ISIT). IEEE, 2022:2255-2260.

[0137]

[11] Fan W, Ma Y, Li Q, et al. Graph neural networks for socialrecommendation[C] / / The world wide web conference. 2019: 417-426.

[0138]

[12] Liu J, Ong G P, Chen X. GraphSAGE-based traffic speed forecastingfor segment network with sparse data[J]. IEEE Transactions on IntelligentTransportation Systems, 2020, 23(3): 1755-1766.

[0139]

[13] Lo WW, Layeghy S, Sarhan M, et al. E-graphsage: A graph neural network based intrusion detection system for IOT[C] / / NOMS 2022-2022 IEEE / IFIP Network Operations and Management Symposium. IEEE, 2022: 1-9.

[0140]

[14] Zhang C, Wang S, Zhan D, et al. Inferring users' social roles with a multi-level graph neural network model[J]. Entropy, 2021, 23(11):1453.

[0141]

[15] Hao Zhifeng, Ke Yanrong, Li Shuo, et al. A social network node classification method based on graph coding network [J]. Computer Applications, 2020, 40(1): 188-195.

[0142]

[16] Khoramnejad F, Joda R, Erol-Kantarci M. Distributed multi-agentlearning for service function chain partial offloading at the edge[C] / / 2021IEEE International Conference on Communications Workshops (ICC Workshops).IEEE, 2021: 1-6.

[0143]

[17] Yang J, Leskovec J. Defining and evaluating network communities based on ground-truth[C] / / Proceedings of the ACM SIGKDD Workshop on MiningData Semantics. 2012: 1-8.

Claims

1. A collaborative caching method based on GNN clustering in mobile edge networks, characterized by: The collaborative caching strategy is used to distribute content to social users in the scenario of social user clustering in the mobile edge network (MEN). The MEN system based on social user clustering includes edge servers, base stations, and terminal social users; Social users are divided into different interest communities based on their social attributes, and different base stations are associated with similar interest communities; the edge server is responsible for caching the request content set of all users in the area. Base station set Responsible for collaborative caching of interest communities The system receives user requests and sends them to community users via multicast. Neighboring base stations exchange information and share content; The edge server is designed with a collaborative caching strategy oriented towards interest communities. The steps for obtaining the collaborative caching strategy include: Step 1: Use the E-GraphSAGE algorithm on social networks to cluster users and establish interest communities with different content preferences; Step 2: Transform the collaborative caching problem for interest communities into a mixed-integer programming problem that maximizes the system's MOS value to meet the QoE requirements of requesting users; use MADDQN to solve the mixed-integer programming problem and obtain the optimal collaborative caching strategy. In step one, 1.1) Preprocessing of social user data The graph structure information contained in the social user dataset includes: ordinary user IDs, celebrity user IDs, and social attributes; In non-Euclidean space, the graph structure of social networks can be used to transform the social user clustering problem into an edge clustering problem: First, construct a bipartite graph from a batch of user information. ,in These represent the sets of ordinary user nodes, celebrity user nodes, and edge sets, respectively; the ordinary user IDs and celebrity user IDs are used as the source nodes. and destination node ; Then, the feature vector of the social attribute The edge features corresponding to the source and destination nodes are used to transform the bipartite graph of the social user dataset into a regular graph. ,in Represents the set of vertices. Represents the set of edges; u and v represent the two endpoints of edge uv, respectively; 1.2) Using E-GraphSAGE to capture graphs Based on the topological structure and edge features, edge embedding and edge clustering are achieved, thus realizing social user clustering; In step two, The mixed integer programming problem is transformed into a Markov decision process (MDP). A multi-agent dual-deep Q-learning network (MADDQN) is used to precisely solve the MDP to obtain the optimal collaborative caching strategy. Specifically: The edge server uses MADDQN to interact with the environment through trial and error, seeking an optimal strategy by maximizing the cumulative reward. ; The MADDQN-based collaborative caching problem can be represented as a triple. ,in Representing the state space, Represents the action space. This represents the immediate reward value gained for completing an action. intelligent agent Before selecting an action, you need to understand the base station. Status information and interest communities under this base station The content requested by the Chinese user Then the state space medium state Represented as ， in ; intelligent agent Received from base station Interest communities within the coverage area Chinese user request content Afterwards, if the local base station Or there is content in the cache of adjacent base stations. Then the action This indicates that no cache placement or update will be performed; if the local base station Or the adjacent base station's cache is empty. Then the action Indicates that the base station Replace a certain content in the cache with the content. ; Environmental conditions Below, intelligent agents Execute action Then proceed to the next state And receive instant rewards. If the optimization goal is to maximize the user's QoE, then the MOS value is set as the reward value. ， where constant Make ; state space To the action space The mapping constitutes a strategy : ; The MADDQN model contains multiple agents, among which... For base stations Determine the collaborative caching strategy for request content across all interest communities within the coverage area; define the current state vector. intelligent agent Use strategy Take action The action-state value function is ， in Indicates base station A collection of all interest communities within the coverage area. Discount factor; According to the Bellman Equation The update formula for the function is: ， in It is the learning rate that controls the learning speed; MADDQN is used to find the optimal collaborative caching strategy, i.e., the agent. Update its main network parameter To approximate the optimal strategy The corresponding Q m value 。 2. The collaborative caching method based on GNN clustering in mobile edge networks according to claim 1, characterized in that: In step two, to maintain the independence of the sampled data, MADDQN establishes an experience replay mechanism, which replays the experience gained by agent m. Experiences are stored in the experience replay pool to record every interaction between the experience and the environment; when the experience replay pool is full, new experiences will randomly replace old experiences. To avoid overestimating the Q-value during training, MADDQN constructs a main network for selecting actions and a target network for estimating the value function, and assumes that the state-action value functions of the two networks are respectively... and ,in This represents the target network parameter.

3. The collaborative caching method based on GNN clustering in mobile edge networks according to claim 1, characterized in that: In step two, the training process of the MADDQN model is as follows: First, agent m randomly samples a small batch of samples from the experience replay pool and inputs them into the main network and the target network, and uses forward propagation to calculate the Q values ​​of the two networks respectively; Next, establish the loss function. ， Backpropagation is performed on the main network to update the main network parameters. Specifically, the gradient of the loss function in equation (18) is calculated. ， Get parameters Update formula ， in Indicates the learning rate; Finally, every Z steps, agent m copies the parameters of the main network to the target network, i.e. Used to update the parameters of the target network. .

Images