A space-air-ground integrated internet relay selection and edge cache update method

Abstract

The application discloses a space-air-ground integrated internet relay selection and edge cache updating method, and belongs to the field of satellite communication. An access relay node with cache capability is established to build a space-air-ground integrated internet access selection and cache configuration model. Based on user historical request data, a graph neural network is used to learn user preference features and content features projected into the same feature space. The user preference features are used for clustering, and the users are clustered into user groups. According to a load balancing heuristic algorithm, the relay node for the user group is selected. According to the preference features and content features of the users served by the relay node, cache configuration updating is carried out. The application mainly aims at the space-air-ground integrated internet scene with relay nodes having cache capability, optimizes the access node selection result of the user, improves the cache hit rate, and thus improves the resource utilization efficiency and user satisfaction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of satellite communication, and particularly relates to a space-air-ground integrated Internet relay selection and edge cache updating method based on a graph neural network. BACKGROUND

[0002] The demand for information by humans has grown tremendously, and in recent years, the demand for short video multimedia files has particularly shown explosive growth. Network transmission content is shifting from traditional services such as text and voice to emerging services that are "content-oriented" streaming media. However, once satellites are deployed, it is difficult to simply and directly expand their spectrum resources and on-board power. Under the condition of limited existing spectrum and maximum on-board power in the satellite segment, how to meet the growing traffic demand of users is one of the core challenges that satellite Internet needs to address.

[0003] Simply relying on optimized configuration of communication resources cannot meet the above challenges. Considering that user requests for multimedia files have high repeatability, which conforms to the 80-20 law, i.e., popular files are repeatedly requested by users. Therefore, deploying edge relay nodes with cache capabilities in the space-air-ground integrated network architecture can effectively avoid repeated transmission of high-frequency files in the space-based information backbone network, improve satellite resource utilization, and significantly reduce end-to-end latency of satellite Internet by avoiding long star-ground links, thereby optimizing user experience.

[0004] (1) For the relay node selection problem of the space-air-ground integrated network, existing methods usually optimize communication performance as the target to make decisions according to factors such as channel quality, distance, signal strength, and load size of each relay node. Such methods are suitable for scenarios where relays do not have cache functions, but for relay nodes with cache capabilities, since users have different content preferences, relay selection and cache hit effect are strongly coupled. Users with similar interests and a tendency to request the same content are connected to the same relay, which can effectively improve cache hit rate and save system transmission resources.

[0005] (2) Existing clustering-based cache strategies often ignore the relay node load balancing problem in communication, and only focus on clustering based on user interest degree, without considering the problem that high load of relay nodes will lead to poor communication performance of the link between users and access relays. SUMMARY

[0006] The application provides a space-air-ground integrated Internet relay selection and edge cache updating method, which is mainly aimed at the scenario of space-air-ground integrated Internet with relay nodes having cache capabilities, optimizes the access node selection result of users, improves cache hit rate, and thereby improves resource utilization efficiency and user satisfaction.

[0007] A space-air-ground integrated Internet relay selection and edge cache update method, characterized in that it comprises the following steps:

[0008] Step one: establish an access relay node with cache capability of space-air-ground integrated Internet access selection and cache configuration model;

[0009] Step two: based on user historical request data, project the user preference features and content features into the same feature space through graph neural network learning;

[0010] Step three: based on user preference features, cluster users into user groups;

[0011] Step four: according to the heuristic algorithm of load balancing, select the access relay node for the user group;

[0012] Step five: according to the preference features and content features of the users served by the relay node, update the cache configuration.

[0013] Preferably, the step one of the application establishes an access relay node with cache capability of space-air-ground integrated Internet access selection and cache configuration model, and the specific process is:

[0014] N U Drones and N T ground relays have cache capability, and the collection of the two is represented as Where L=N U +N T ; the directory of all N contents is represented as Assuming that the data volume of all contents is the same S F , each relay unit has a M×S F bit memory, which can store up to M contents; K users are distributed in the network coverage, represented by

[0015] The established relay selection and cache configuration optimization problem is:

[0016]

[0017] Where the optimization variables C t and X t represent the cache decision matrix of each relay node and the relay access selection matrix of each user at time t, respectively, and the two optimization objectives and represent maximizing user latency satisfaction and resource saving rate of the satellite segment, respectively, where US t represents the user latency satisfaction at time t, and SR t represents the resource saving rate at time t. In the first constraint condition, ​represents the user u k whether to request content f n The constraint indicates that the user can request at most one content in a time slice; in the second constraint condition, x k,l represents the result of the user accessing the relay, i.e. x k,l = 1 indicates that the user u k accesses the relay node R l The constraint indicates that the user can only access one relay node for data request at a certain moment; in the last constraint, R QoS represents the minimum transmission rate requirement, and the constraint indicates that the transmission rate between the user and the relay node needs to meet the minimum transmission rate requirement, and if the constraint cannot be met, the content request of the user will be rejected by the relay node.

[0018] Preferably, in step two of the present application, based on the historical request data of the user, the user preference features and the content features projected into the same feature space are learned by a graph neural network, and the specific process is as follows:

[0019] Step 2.1 represents a pair of training data, wherein f and j both represent contents, f represents the content requested by the user u, and j represents the content not requested by the user u, represents a set of user and content pairs observed from the user access history, represents a set of user and content pairs without content access records;

[0020] Step 2.2 uses a graph neural network to realize first-order propagation, including message construction and message aggregation processes;

[0021] Step 2.3 uses a graph neural network to realize high-order propagation on the basis of first-order propagation;

[0022] Step 2.4 uses the user-content pair training data to perform supervised training on the graph neural network.

[0023] Preferably, in step three of the present application, the users are clustered into user groups based on the user preference features, and the specific process is as follows:

[0024] Step 3.1 uses a graph neural network to obtain the features of each user

[0025] Step 3.2 calculates the preference similarity matrix Sim between users using cosine similarity i,j ]

[0026]

[0027] wherein and respectively represent the user features obtained by using the graph neural network in step 3.1;

[0028] Step 3.3 obtains G disjoint user clusters by performing K-means clustering on the user preference similarity matrix Sim where the interest similar nodes should be in the same cluster.

[0029] Preferably, in step four of the present application, a heuristic algorithm based on load balancing is used to select the relay nodes accessed by the user group, and the specific process is as follows:

[0030] Step 4.1 Relay node R l The load of the relay node R where x i,l represents the relay access result of the i-th user cluster, represents the number of users in the i-th user cluster, and the load balancing coefficient is represented as

[0031] Step 4.2 Since the users accessing the same relay node share their spectrum resources, the user cluster access relay needs to maintain the load balancing between relay nodes, so the access node selection problem of the user cluster is modeled as:

[0032]

[0033] Step 4.3 proposes a two-stage heuristic access selection method, which directly selects the access of the user cluster, thereby obtaining the user relay access selection result considering load balancing and content preference.

[0034] Preferably, in step five of the present application, the cache configuration is updated according to the preference features and content features of the users served by the relay nodes; the specific steps are as follows:

[0035] Step 5.1 The cloud data center regularly updates the graph neural network parameters using historical request data, and calculates the preference features of each user

[0036] Step 5.2 The cloud data center calculates the average preference features of each relay access user, and for the relay node R l , the average preference features of the users are:

[0037]

[0038] where represents the preference features of the user u, which are obtained by the graph neural network; and the average preference features are distributed to each relay node by the satellite;

[0039] Step 5.3. The content requested by the user is judged. If the content is in the cache of the relay node, the relay node directly transmits the content to the user. If the content is not in the cache, each relay node respectively counts the content requests in each time slice that are not in the local cache, and periodically reports the statistical data to the satellite;

[0040] Step 5.4. For the above missed content, the satellite backhaul link obtains the content f and the features of the content from the cloud data center , and then transmits f and to all relay nodes through broadcast;

[0041] Step 5.5. Each relay node calculates the average interest of the connected users in the content f, for example, for the relay node R l , the average interest is:

[0042]

[0043] where x k,l = 1 represents that the user u k selects the relay node R l to access, The interest of the user u in the content f is represented by the inner product operation, and the interest of the user in the content is obtained.

[0044] Step 5.6. Each relay node calculates the average user interest of each content that has been cached, and uses the average user interest as an indicator to determine whether to replace the existing content in the cache with the current content.

[0045] The technical scheme of the present application has the following advantages compared with the prior art:

[0046] (1) Considering the influence of relay selection on cache hit rate in the space-ground-integrated Internet scenario with relay caching capability, a relay selection and cache content placement joint optimization model is established to improve user satisfaction and save satellite resources.

[0047] (2) According to the historical data of user request content, the projection of the user and the content in the same feature space is obtained by using the graph neural network, and the user is clustered by using the preference feature of the user to the content, so that users with similar interests tend to access the same relay, and the relay node can cache a small number of content interested by such users with limited cache space, without needing to focus on the overall popular content among all users, which helps to improve the cache hit rate of the relay under the condition of limited relay cache capacity, thereby optimizing the resource utilization rate and user delay satisfaction.

[0048] (3) In the process of solving the user access selection problem, not only the user's content preference is considered, but also the load balancing problem in the communication field. A heuristic load balancing access method is proposed to solve the access selection problem of user clusters, so that the user access selection scheme takes into account both the user preference of caching and the load balancing of relay nodes in communication.

[0049] (4) Using graph neural networks to perform collaborative filtering on historical data of user-requested content, compared with existing matrix factorization-based collaborative filtering methods, it is more suitable for large-scale users and content in actual Internet caching problems; and it has good scalability, making it easy to dynamically fine-tune the graph neural network as user content requests continue and user interests change, thereby updating user interest features. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the application scenario of the present invention, namely, an integrated air-space-ground internet scenario with relay caching capabilities.

[0051] Figure 2 This is a schematic diagram illustrating the service process of whether the edge node matches the content.

[0052] Figure 3 This is a schematic diagram of the relay node selection and cache content update time scale in this invention;

[0053] Figure 4 This is a flowchart of the node relay selection process in this invention;

[0054] Figure 5 It is a user content bipartite graph and a schematic diagram based on its higher-order connectivity. Detailed Implementation

[0055] The present invention will be further explained below with reference to the accompanying drawings and specific implementation process.

[0056] I. Establish an integrated air-space-ground internet access selection and caching configuration model with caching capabilities for access relay nodes;

[0057] This invention focuses on the downlink of a content-centric, integrated air-space-ground internet scenario, such as... Figure 1 As shown. Assuming users cannot communicate directly with the satellite, content from the internet server is relayed by the LEO satellite to a drone or ground station before being transmitted to the user. The system operation can be discretized into a time series containing equal time slices.

[0058] N U One drone and N T The set of ground relays is represented as Where L = N U +NT , are equipped with a cache unit with limited capacity. The catalog of all N contents is denoted as For ease of discussion, assume that the data size of all contents is S F . Each relay's cache unit has a M x S F memory, which can cache up to M contents. K users are distributed in the network coverage, denoted as , assuming that users are static.

[0059] In the integrated space-ground-ground network scenario, the data request of a ground user terminal can be satisfied in two ways, as shown in Figure 2

[0060] (1) Figure 2 (a) is the ideal case, when the requested content is cached locally in the accessed relay, the local relay directly transmits the file to the user. Even so, this request will be recorded for subsequent relay node cache updates.

[0061] (2) Figure 2 (b) represents when the local relay node cache does not have the requested content, the content is obtained from the cloud data center by the LEO constellation. The specific process is that each relay node respectively counts the content requests that are not cached locally in each time slice, and periodically reports the statistics to the satellite. For the above missed contents, the satellite backhaul link obtains the contents from the cloud data center, and then transmits these files to the relay nodes through broadcast, and the relay nodes distribute the judgment whether to cache the current satellite broadcast push content.

[0062] This satellite broadcast-based content distribution mechanism takes advantage of the wide-area coverage of satellites and can realize a combination of active and passive caching mechanisms, which is beneficial to improve the cache update efficiency. For example, the user content request f0 of relay node R0 misses, the satellite obtains f0 from the cloud data center and broadcasts it to all relay nodes. All relay nodes receive the content f0 and judge whether to cache it, so that the relay R1 can pre-cache this content before the local user requests f0.

[0063] Let matrix X = [x k,l ], represent the relay selection results of all users, where x k,l ∈{0,1}, x k,l = 1 indicates that user u k selects to access relay node R l . Users can only access one relay node, while relay nodes can serve multiple users.

[0064] B represents the system bandwidth, and Pt l is the transmission power of the relay node R l ​Average transmission power, For user u k To relay R l The distance between them. The channel quality parameters between the user and the relay node are... The channel gain can then be expressed as in and These represent the transmit and receive antenna gains, respectively.

[0065] Therefore, user u k With relay node R l The signal-to-interference-plus-noise ratio (SIR) of the downlink can be expressed as:

[0066]

[0067] Where Pt q Indicates the transmission power of other relay nodes. R represents the downlink channel gain of other relay nodes. The relay distributes bandwidth evenly among the accessing users, therefore R... l With user u k The maximum downlink transmission rate between them can be expressed as:

[0068]

[0069] For transmitting certain content, the transmission delay of this link segment can be expressed as:

[0070]

[0071] This invention does not discuss the routing problem of the LEO constellation; therefore, it is assumed that the round-trip latency for the LEO satellite constellation to access the cloud service center and obtain data is a fixed value, denoted as:

[0072] Satellite to relay R l ≠ Downlink channel gain for data transmission Among them G S This refers to the transmission gain of the satellite antenna. These are channel quality parameters, including large-scale fading and small-scale fading. For relay R l The receiving antenna gain. Satellite and relay R l The downlink signal-to-noise ratio between them can be expressed as:

[0073]

[0074] Where P S For satellite transmission power, For relay node R l Noise parameters.

[0075] The maximum transmission rate of each relay node is

[0076]

[0077] Therefore, the link transmission delay can be expressed as

[0078]

[0079] Assume that each user requests at most one content at each time, let the vector represent the content request of user u k At time t, the content request is

[0080]

[0081] It is generally believed that the request probability of each content follows Zipf distribution, and its popularity changes over time. If the content f n is the r n th most frequently accessed content in all contents, then its average request probability can be approximately expressed as

[0082]

[0083] where γ is the decay constant describing the popularity distribution of the content. The larger γ indicates that the file popularity distribution is more uneven.

[0084] On the basis of the temporal change of popularity, different users have different preferences for content, which can be expressed as

[0085]

[0086] Let the matrix C represent the placement results of each relay node's cache, i.e.

[0087]

[0088] Therefore, whether the file accessed by user u k is in the cache of the access relay node can be represented by , where the vector represents the access relay selection result of user u k .

[0089] If , the requested data of user u k is in the cache of the relay node, and the relay node can directly transmit the requested data of the user, and the transmission delay is

[0090]

[0091] like This requires accessing the cloud server via satellite to obtain the content, and then forwarding the content to the user terminal through a relay node. The transmission latency is:

[0092]

[0093] in This is the return time required for content to be sent from the cloud server to the satellite, and is a fixed value.

[0094] User u at time t k The end-to-end latency of the requested content can be expressed as:

[0095]

[0096] User u k The content f requested at time t n The latency satisfaction index can be expressed as:

[0097]

[0098] Where D Max This represents the maximum acceptable latency for users in the system.

[0099] Therefore, the average satisfaction of all users in the system with the latency index can be expressed as:

[0100]

[0101] The amount of content transmitted by the system to users within the t-th time slice is the sum of the amount of content data requested by each user through the system, i.e.

[0102]

[0103] The satellite needs to transmit content that is not cached by the local relay node. The amount of data to be transmitted in the t-th time slice is:

[0104]

[0105] Where S F For the amount of data in a single piece of content, Indicates user u k The content requested at time t is in the relay of the user access, so it does not need to be obtained from the cloud data center using the space-based segment.

[0106] Therefore, the savings rate of satellite segment resource overhead can be expressed as:

[0107] It is equivalent to the overall system cache hit rate.

[0108] In summary, the optimization model corresponding to the relay selection and cache strategy problem of the patent can be represented as follows:

[0109]

[0110] The optimization variables C in the optimization problem (19) t and X t respectively represent the cache decision matrix of each relay node and the relay access selection matrix of each user at time t, and the two optimization objectives and respectively represent maximizing user latency satisfaction and resource cost saving rate of satellite segment, and the three constraint conditions, the first two respectively represent that a user can request at most one content in a time slice and a user can only access one relay node for data request at a certain time; the last constraint indicates that the transmission rate of the user and the relay node needs to meet the minimum transmission rate requirement, otherwise the user's content request will be rejected by the relay node.

[0111] The time scales of the two decision problems of relay selection and cache content update are different, as shown in Figure 3 Due to the signaling overhead of relay switching and the stability of user interest degree in a long time range, the relay selection of each user should be stable in a long time to avoid frequent switching; the replacement overhead of local cache is small, and the dynamic changes of content popularity and user preference can be tracked by frequently updating the cache content. Therefore, the time scales of the two decision problems of relay selection and cache content update are different.

[0112] Due to the different interest degrees of different users for different contents, the user access based on interest degree can effectively optimize the cache hit rate on the premise of determining the cache strategy, thereby reducing the average latency of users. However, the above work only considers the improvement of cache performance, and does not consider the load balancing of access nodes required for communication performance optimization. Therefore, the inventor believes that in the scenario where the relay node has cache capability, the user access selection problem should consider both cache and communication effects, and proposes a user access selection algorithm based on graph neural network user interest distance clustering and load balancing, as shown in Figure 4 .

[0113] Two: based on the historical request data of users, the user preference features and content features projected into the same feature space are learned by graph neural network; the specific process is as follows:

[0114] Assume that the system has been running for a period of time and has accumulated historical request records of each user, denoted as {<u k ,f n ,t reg >}, where each element <u k ,f n ,treg represents that at t reg moment, user u k requests content f n .

[0115] Based on the user history request record, the user and content feature learning is carried out by using an inductive graph neural network method, and the core idea is that the feature of a node in a graph can be obtained by aggregating the features of adjacent nodes, and the feature embedding of the node can be obtained by learning accurate feature aggregation parameters. Therefore, the propagation embedding method is adopted, the high-order connectivity in the user and content bipartite graph is utilized, the historical request record of the user to the content is explicitly used as a supervision signal, and the projection of the user and the content in the same feature space is learned.

[0116] The idea is shown in Figure 5 , the request matrix of the user to the content is converted into a bipartite graph form, and the graph neural network is used to perform step-by-step feature projection according to direct connection, indirect connection and cross-layer connection. Figure 5 The left side shows that the existence of the edge between the user node and the content node in the content-user request relationship bipartite graph indicates that the user has requested the content. Figure 5 The right side represents a three-layer connection graph structure of the user node. For example, for user 1, the first layer is the content it has accessed, including content 1, content 2 and content 4; the second layer is the user who has requested the content accessed by user 1, including user 4 and user 6; the third layer is the content requested by the user in the second layer, including content 3 and content 6. Therefore, the content and user features in the three layers are considered, and the feature aggregation is performed based on the graph structure, so that the feature of user 1 can be obtained. The patent adopts a graph neural network to learn the parameters in the aggregation operation.

[0117] 1, first-order propagation

[0118] The request of the user to the content provides a direct basis for the user preference, so the content feature requested by the user can be regarded as the content feature basis for measuring the similarity of two users. On this basis, feature propagation can be performed between the user and the content with a request relationship. This propagation is completed through two main operations, which are message construction and message aggregation.

[0119] The message construction adopts a feature combination function m u←f =f(e f ,e u ,p ui ) to construct information from the content requested by the user to the user, wherein e f and e u represent the initial embedding vectors of the content and the user, respectively, which can be constructed in the form of a word vector according to the general practice in the recommendation system. This function utilizes the content feature requested by the user, the user feature and the parameter p u,fEncoding message features from the content requested by the user is implemented.

[0120] Considering the graph relationship, the message combination function is defined as

[0121]

[0122] where W1, represents the feature projection parameter matrix of the neural network that needs to be trained, which acts on extracting useful information and propagating, d is the feature dimension of the content and the user feature e f and e u before the message construction operation, and d' is the feature dimension of the message after the message construction. The symbol ⊙ represents the multiplication of the corresponding positions of the feature vectors e f and e u , which helps to encode more content features associated with the user into the propagated information.

[0123]

[0124] is the parameter representing the relationship between the user and the content in the message construction process, where and respectively represent the number of neighbor users or neighbor contents directly associated with the current user or content, and the parameter p u,f can be understood from the aspect of representation learning, that is, the parameter p u,f represents the contribution degree of the content accessed by the user to the user interest feature vector, so the more the user accesses the content, the larger the number of content neighbors on the access graph , and the smaller the contribution of each content to its interest degree.

[0125] In the message aggregation stage, the information transmitted by each neighbor of the user needs to be aggregated to obtain the user feature. The aggregation method is

[0126]

[0127] where represents the user feature obtained by the user through the first-order propagation mechanism; represents the aggregation of the messages propagated from each accessed content by using the direct summation method, that is, the equality between the requested contents; on the basis of the aggregation term , the user's own feature m u←u is superimposed, which helps to maintain its original features, and the activation function LeakyReLU allows the feature to be positive or a small negative number.

[0128] The above message construction and message aggregation process focuses on the process by which user nodes obtain feature mappings through first-order propagation. For content nodes, the principle is the same, except that the user requesting the content is taken as its neighbor node.

[0129] 2. Higher-order communication

[0130] The first-order propagation mechanism described above can aggregate user features using the content features directly accessed by the user, or obtain aggregated content features using users who directly access the content. Therefore, the first-order connectivity model can achieve a projection of user and content features based on direct access records.

[0131] Building upon first-order propagation, more layers of embedded feature propagation can be stacked to optimize feature learning by leveraging higher-order connectivity information. In other words, user interest features are not only related to the content they directly access, but also need to consider the interest features of users who access the same content and the content these users have accessed.

[0132] By stacking embedded propagation layers, users (and projects) are able to receive messages propagated from their l-hop neighbors. For example... Figure 5 As shown, in step l, the features of user u can be recursively represented as:

[0133]

[0134] Its message construction process can be represented as follows:

[0135]

[0136] The parameters of each propagation layer All of these require learning, and the learning methods are introduced below. For example... Figure 5 As shown on the right, through the propagation path of the feature information from content 6 → user 6 → content 4 → user 1, the message from content 6 is encoded in... Therefore, stacking multiple embedding propagation layers can effectively inject higher-level user-content association information into both user and content feature learning.

[0137] 3. Interest estimation

[0138] Through first-order and higher-order propagation mechanisms, based on known user and content access records, it is possible to effectively obtain user and content feature projections from user access records, namely: and By concatenating the features output from each propagation layer, the complete features of the user and content can be obtained, i.e.:

[0139]

[0140] Since the feature learning has been done by using the graph neural network, the user and content features are projected into the same feature space, so the user's interest in the content can be obtained by using a simple inner product operation, i.e.

[0141]

[0142] where represents the interest degree of the user u to the content f.

[0143] 4. Graph neural network training

[0144] The key to the projection effect of the user and content features lies in learning the propagation layer parameters of each propagation layer by the graph neural network

[0145] The BPR loss commonly used in recommendation systems and collaborative filtering is used to learn the propagation layer parameters of the network, and the BPR assumes that the observed interaction can better reflect the user's preference, and should be assigned a higher predicted value than the unobserved interaction, which is calculated as follows:

[0146]

[0147] where represents a pair of training data, represents a set of user and content pairs for which the user access history is observed, represents a set of user and content pairs for which no content access record is available. and respectively represent the associated interest degree values between the user u and the content f and the content j after the feature projection based on the graph neural network. represents all the parameters to be learned in the network, where W e is the parameter matrix for projecting the word vector obtained from the user and content ID to the feature vector and .

[0148] Three: based on the user preference feature, the users are clustered into user groups; the specific process is as follows:

[0149] The purpose of the user clustering task is to divide the users into G disjoint user clusters according to the interest degree feature where the interest similar nodes should be in the same cluster.

[0150] The user interest features are obtained by using the graph convolutional neural network The cosine distance is used to measure the interest distance between users. Since the features have been well represented before, a relatively simple and basic Kmeans method is used to cluster them, which can obtain a good clustering effect.

[0151] Four: according to the load balancing heuristic algorithm, the access relay node of the user group is selected; the specific process is as follows:

[0152] Let denote the number of users in the i-th user cluster, denote the selection result of the user group accessing the relay node, x i,l =1 indicates that the user cluster g i accesses the relay node R l , and a cluster of users can only access one relay node. By clustering, the access selection problem of users to relays is converted into the access selection problem of user clusters with similar interest degrees to relays.

[0153] The load of the relay node R l may be represented by the number of users accessing it, that is:

[0154]

[0155] Therefore, the load balancing coefficient can be represented as:

[0156]

[0157] Since users accessing the same relay node share its spectrum resources, the user cluster accessing the relay needs to maintain the load balance among relay nodes, so the access node selection problem of the user cluster can be modeled as:

[0158]

[0159] The present application proposes a two-stage heuristic access selection method, which simply and directly selects the access of the user cluster, as shown in Algorithm 1.

[0160]

[0161]

[0162] Five: according to the preference characteristics and content characteristics of the users served by the relay node, the cache configuration is updated, and the specific process is as follows: step 5.1: the cloud data center regularly updates the graph neural network parameters using historical request data, and calculates the preference characteristics of each user Step 5.2: The cloud data center calculates the average preference characteristics of each relay access user, for example, for the relay node R l , the average preference characteristics of its users are:

[0163]

[0164] Where The preference feature of the user u is obtained by a graph neural network, and the average preference feature is distributed to each relay node by a satellite.

[0165] Step 5.3: judging the content requested by the user, if in the cache of the relay node, the relay node directly transmits the content to the user; if not in the cache, each relay node respectively counts the content requests not in the local cache in each time slice, and regularly reports the statistical data to the satellite.

[0166] Step 5.4: for the above missed content, the satellite backhaul link obtains the content f and the feature of the content from the cloud data center , and then transmits f and to all relay nodes by broadcasting.

[0167] Step 5.5: each relay node calculates the average interest degree of the user to the content f, for example, for the relay node R l , the average interest degree is:

[0168]

[0169] wherein x k,l = 1 represents that the user u k selects the relay node R l to access, The interest degree of the user u to the content f is represented, since the feature learning has been performed by the graph neural network, the user and the content feature are projected into the same feature space, and therefore, the interest degree of the user to the content can be obtained by using only a simple inner product operation.

[0170] Step 5.5: each relay node calculates the average user interest degree of each content already cached, and judges whether the current content is replaced with the existing content in the cache according to the average user interest degree. The specific process is shown in Algorithm 2

[0171]

[0172]

[0173] The above only describes the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacements or changes to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, without departing from the spirit and scope of the present application, which should be covered within the protection scope of the present application.

Claims

1.A method for relay selection and edge cache update in space-air-ground integrated Internet, characterized in that Comprise the following steps: Step one: establish access relay node has the cache ability space-air-ground integrated Internet access selection and cache configuration model; The specific process is: N U One drone and N T Each ground relay has caching capabilities, and the set of both is represented as: Where L = N U +N T A directory containing all N items is represented as follows: Assume all content has the same data volume, S. F Each relay's cache unit has an M×S F A 1-bit memory can cache up to M items; K users are distributed across the network coverage area, using... express; The established relay selection and cache configuration optimization problem is: where C is the optimization variable t and X t represent the buffer decision matrix of each relay node and the relay access selection matrix of each user at time t respectively, two optimization objectives and maximize the user latency satisfaction and the resource saving ratio of satellite segment respectively, where US t denotes the user latency satisfaction at time t, SR t denotes the resource saving ratio at time t; in the first constraint condition, denotes whether user u k requests content f n , this constraint indicates that a user can request at most one content in a time slice; in the second constraint condition, x k,l denotes the result of user access to relay, i.e. x k,l = 1 indicates that user u k accesses relay node R l , this constraint indicates that a user can access at most one relay node for data request at a time; in the last constraint condition, denotes the maximum downlink transmission rate between relay node R l and user u k , R QoS denotes the minimum transmission rate requirement, this constraint indicates that the transmission rate between user and relay node needs to meet the minimum transmission rate requirement, if this constraint cannot be met, the content request of the user will be rejected by the relay node; Step two: based on the user historical request data, through the graph neural network learning projection to the same feature space of user preference feature and content feature; Step three: based on user preference feature clustering, the user is clustered into user group; Step four: according to the heuristic algorithm of load balancing, the access of relay node is selected for user group; Step five: according to the preference feature and content feature of the user served by the relay node, the cache configuration is updated. 2.The method of claim 1, wherein The above step two based on user historical request data, through the graph neural network learning projection to the same feature space of user preference feature and content feature, the specific process is: Step 2.1 represents a pair of training data, where f and j both represent content, f represents content requested by user u, and j represents content not requested by user u, represents a set of user and content pairs for which user access history is observed, represents a set of user and content pairs for which there is no content access record. Step 2.2 uses graph neural network to realize first-order propagation, including message construction and message aggregation process; Step 2.3 on the basis of first-order propagation, using graph neural network for high-order propagation; Step 2.4 Training data by user-content pairs Supervised training of the graph neural network. 3.The method of claim 2, wherein The above step three based on user preference feature clustering, the user is clustered into user group, the specific process is Step 3.1 Obtain features of each user by using graph neural network Step 3.2 Calculate the preference similarity matrix Sim = [sim i,j ], between users using cosine similarity. Step 4.2 Calculate the similarity matrix Sim = [sim i,j ] between users using cosine similarity. Step 5.2 Calculate the similarity matrix Sim = [sim i,j ] between users using wherein and respectively represent the user features obtained by step 3.1 using a graph neural network; Step 3.3 Get G disjoint user clusters by K-means clustering on the preference similarity matrix Sim of users where the interest similar nodes should be in the same cluster. 4.The method of claim 3, wherein The above step four according to the heuristic algorithm of load balancing, the access of relay node is selected for user group, the specific process is: Step 4.1 Relay node R l The load of the relay node R is represented as the number of its access users, i.e. where x i,l represents the relay access result of the i-th user cluster, represents the number of users in the i-th user cluster, and the load balancing coefficient is represented as Step 4.2 because the user accesses the same relay node shares its spectrum resources, so the user cluster access relay needs to keep the load balance between relay nodes, so the access node selection problem of user cluster is modeled as: Step 4.3 a two-stage heuristic access selection method is proposed, which directly selects the access of user cluster, so as to obtain the user relay access selection result considering load balancing and content preference. 5.The method of claim 4, wherein The above step five according to the preference feature and content feature of the user served by the relay node, the cache configuration is updated; The specific steps are as follows: Step 5.1 The cloud data center periodically updates the graph neural network parameters with historical request data and computes the preference features for each user Step 5.2 The cloud data center computes the average preference profile of each relayed access user, for the relay node R l with average preference profile of its users as: wherein represents the preference feature of the user u, obtained by the graph neural network; the average preference feature pref of each relay node is distributed to each relay by the satellite l is distributed to each relay Step 5.3 judges the content requested by the user, if in the cache of relay node, the relay node directly transmits the content to the user; If not in the cache, each relay node respectively counts the content request which is not in the local cache in each time slice, and reports the statistical data to the satellite regularly; Step 5.4 For the above missed content, the satellite backhaul link fetches the content f and the features of the content from the cloud data center and f are then broadcast to all relay nodes transmit to all relay nodes; Step 5.5 Each relay node computes the average interest of the connected users in the content f, for the relay node R l , the average interest is: where x k,l = 1 represents a user u k selecting a relay node R l accessing, characterizing the interest of a user u in a content f, using an inner product operation, obtaining the interest of the user in the content; Step 5.6 Each relay node calculates the average user interest of each content that has been cached, and judges whether to replace the existing content in the cache with the current content according to the average user interest. Step 5.6 Each relay node calculates the average user interest of each content that has been cached, and judges whether to replace the existing content in the cache with the current content according to the average user interest.

Images