Satellite network semantic dual-channel multi-demand access recommendation method

Abstract

This invention discloses a dual-channel, multi-demand access recommendation method for satellite networks, belonging to the field of satellite communication network resource management technology. It acquires historical access resource interaction sequences of terminals, integrates resource identifier embedding and semantic attribute embedding, and extracts multiple communication demand representations through time-series coding and routing. It calculates the mixed similarity of communication demands and semantic demands, constructs a collaborative relationship graph and a competitive relationship graph based on high- and low-similarity neighbors, and obtains refined demand representations through dual-channel graph propagation and residual fusion. It filters candidate resources based on satellite visibility, beam coverage, and basic service capabilities, calculates recommendation scores by combining multi-demand matching, attention aggregation, and short-term context, and outputs the recommended access resources for the next time step. During training, the model parameters are optimized using real next-step interactive access resources as supervision signals. This invention integrates historical behavior and resource semantics, explicitly models multi-demand collaborative and competitive relationships, and is suitable for dynamic satellite network resource recommendation.

Description

Technical Field

[0001] This invention relates to the field of satellite network access control and intelligent information processing technology, specifically a method for recommending semantic dual-channel multi-demand access to satellite networks. Background Technology

[0002] Terminal service requests in satellite networks are characterized by multiple demands, significant differences in service categories and quality of service requirements, and changes in satellite and beam visibility over time. Terminals interact with satellites, beams, gateways, or satellite network service resources continuously. Access recommendation typically obtains the terminal's historical interaction sequence, resource identifiers, and resource attributes, and uses this to form resource and demand representations. Candidate resources are then selected from the current resources and ranked according to their matching degree to complete the access resource selection for the next moment.

[0003] Given that satellite and beam visibility vary over time and terminals have multiple parallel service demands, existing methods rely on a single relationship graph or shared propagation parameters to express demand relationships. They often directly use historical matching results of candidate resources for ranking, making it difficult to simultaneously separate collaborative and competitive relationships between demands. Furthermore, it's challenging to transfer constraints on current visibility, beam coverage, and basic service capabilities to the candidate set. Simultaneously, single behavioral relationships are difficult to characterize demand relationships in conjunction with resource semantics, and the short-term context formed by the most recent interaction is not stably used for ranking. This results in difficulties in separating the propagation of collaborative enhancement and competition suppression, insufficient constraints on the current network executability of candidate resources, inadequate joint expression of behavioral relationships and resource semantics, and insufficient utilization of the immediate access intent of the short-term context. Therefore, a dual-channel, multi-demand access recommendation method based on satellite network semantics is needed to address the aforementioned problems of separating the propagation of collaborative enhancement and competition suppression, constraining the current network executability of candidate resources, jointly expressing behavioral relationships and resource semantics, and stably utilizing the short-term context. Summary of the Invention

[0004] In view of the above-mentioned problems in the prior art, the present invention provides the following technical solution: This invention provides a satellite network semantic dual-channel multi-demand access recommendation method, including: Obtain the historical access resource interaction sequence of the terminal sorted by time, integrate the identifier embedding and semantic attribute embedding of each access resource, and superimpose the position embedding to obtain the sequence representation; The sequence representation is subjected to time-series encoding and multi-demand routing to obtain the communication demand representation; Calculate the collaborative similarity between the communication requirement representations, extract semantic requirement representations using the semantic attribute embedding and calculate semantic similarity, and weightedly fuse the two similarities to form a hybrid similarity matrix; Using the communication requirements as nodes, a collaborative relationship graph and a competitive relationship graph are constructed according to the high and low similarity in the hybrid similarity matrix, respectively. Based on the cooperative relationship graph and the competitive relationship graph, the communication requirement representation is propagated to obtain the cooperative representation and the competitive representation, and then fused with the residual of the communication requirement representation to obtain the refined requirement representation; Candidate access resources that meet the conditions of terminal visibility, beam coverage and basic service capabilities are selected. The matching score between each candidate access resource and the refined requirement representation is calculated. The recommendation score is obtained through requirement attention aggregation. The resources are sorted according to the recommendation score and the access recommendation result for the next time step is output.

[0005] Optionally, the access resources are satellites, beams, gateways, satellite network service resources, or resource units formed by a combination of satellite identifiers, beam identifiers, and service attributes; The semantic attribute embedding is obtained by encoding at least one of the following attributes: resource type, service description, business category, bandwidth, latency, reliability, coverage area, and service capability. Continuous attributes are discretized, labeled, or jointly encoded with text attributes before participating in the encoding. The identifier embedding and semantic attribute embedding are achieved by mapping the semantic attribute embedding to the identifier embedding space through a semantic projection matrix, and then multiplying the projection result by a semantic fusion coefficient and adding it to the identifier embedding.

[0006] Optionally, the sequence representation process includes: The sequence representation is encoded using a Transformer encoder with causal masking, so that each interaction position uses only the information of the current interaction position and the previous positions; The encoding results are softly assigned to a preset number of demand components through a routing mechanism, and then aggregated to obtain the communication demand representation.

[0007] Optionally, the collaborative similarity is the cosine similarity between two communication requirement representations; The semantic requirement representation is extracted through a pure semantic sequence, which is taken from the semantic attribute embedding and processed by Transformer encoding and routing. The semantic similarity is the cosine similarity between two semantic requirement representations. Each element in the hybrid similarity matrix is ​​obtained by weighted sum of the collaborative similarity and semantic similarity of the same demand node pair, with the weights of the two similarities being the hybrid similarity coefficient and one minus the hybrid similarity coefficient, respectively.

[0008] Optionally, in the collaborative relationship graph, a preset number of neighbors with the highest mixed similarity are retained for each demand node, and the edge weight is obtained by dividing the mixed similarity by the positive temperature parameter, taking the exponent, and normalizing it among the retained neighbors; In the competitive relationship graph, a preset number of neighbors with the lowest mixed similarity are retained for each demand node. The edge weight is obtained by dividing the negative number of the mixed similarity by the positive temperature parameter, taking the exponent, and normalizing it among the retained neighbors. The edge weight corresponding to the unretained neighbors is zero.

[0009] Optionally, the formation of the cooperative and competitive representations includes: The required node information is aggregated based on the adjacency matrix of the collaborative relationship graph, and the collaborative representation is obtained through the positive parameter matrix and ReLU activation. The demand node information is aggregated based on the adjacency matrix of the competition relationship graph, and the competition representation is obtained through the negative parameter matrix and ReLU activation. The positive parameter matrix and the negative parameter matrix are not shared, forming the cooperative representation and the competition representation.

[0010] Optionally, the formation of the refinement requirement representation includes: The communication requirements are represented as the residual baseline; The refined requirement representation is obtained by multiplying the collaborative representation by the positive enhancement coefficient and adding it to the residual benchmark, and then subtracting the product of the competition representation and the negative suppression coefficient.

[0011] Optionally, the candidate access resources are filtered based on the visibility, coverage, and service capability information provided by the satellite ephemeris or network control plane at the current recommended time, excluding access resources that do not meet any of the terminal visibility, beam coverage, and basic service capability conditions.

[0012] Optionally, the determination of the recommendation score includes: The matching score is obtained by calculating the inner product between each refined requirement representation and the resource representation of the candidate access resource; For each matching score of the same candidate access resource, a softmax operation is performed to obtain the demand attention weight, and the matching scores are weighted and aggregated according to the demand attention weight. The encoding state corresponding to the most recent interaction is used as the short-term context representation. The inner product between the short-term context representation and the resource representation of the candidate access resource is calculated to obtain the context score. The weighted aggregation result is added to the context score, and the resource bias of the candidate access resource is added to obtain the recommendation score.

[0013] Optionally, the parameters for calculating the recommendation score are obtained by training to minimize the joint loss. The joint loss is formed by weighted summation of the recommendation loss with the real next-interaction access resource as the supervision signal, the demand diversity loss to reduce redundancy between different communication demand representations, and the cooperative-semantic alignment loss to improve the consistency between access behavior representation and resource semantic representation.

[0014] The technical solution provided by this invention has the following beneficial effects: 1. By fusing identifier embedding and semantic attribute embedding, and constructing collaborative similarity and semantic similarity respectively, historical interaction relationships and resource service semantics jointly participate in demand relationship calculation, forming a hybrid similarity matrix, reducing the lack of demand relationship expression caused by relying solely on resource identifiers or single behavioral relationships.

[0015] 2. A collaborative relationship graph is constructed by using highly similar neighbors and a competitive relationship graph is constructed by using low-similarity neighbors. A dual-channel propagation without shared parameters and residual fusion with coefficients in opposite directions are adopted so that the collaborative representation and the competitive representation can respectively play the roles of enhancement and suppression. At the same time, the residual path of the communication demand representation is preserved, reducing the loss of original demand information when the neighbor relationship is insufficient.

[0016] 3. The system filters currently executable candidate access resources by visibility, beam coverage, and basic service capabilities. Then, it combines demand attention aggregation, short-term context score, and resource bias to determine the recommendation score, so that the ranking process is simultaneously constrained by multi-demand matching, the most recent interaction state, and the current network executability. Through joint training of recommendation loss, demand diversity loss, and collaborative-semantic alignment loss, the system reduces the redundancy of multi-demand representations and maintains the consistency between behavioral representations and resource semantics. Attached Figure Description

[0017] Figure 1 A step diagram of a satellite network semantic dual-channel multi-demand access recommendation method provided by the present invention; Figure 2 The present invention provides an overall flowchart of a satellite network semantic dual-channel multi-demand access recommendation method; Figure 3 This is a Chinese schematic diagram of the semantic dual-channel relationship graph model provided by the present invention; Figure 4 This is a schematic diagram of the multi-demand dual-channel relationship matrix provided by the present invention, wherein, Figure 4 (a) is the cooperative relationship matrix. Figure 4 (b) is the competition relationship matrix. Figure 4 (c) is a two-channel relationship difference matrix. The color intensity in the figure is used to indicate the strength of the relationship or the direction of the relationship difference. Detailed Implementation

[0018] The specific embodiments of the present invention will now be described with reference to the accompanying drawings. The described embodiments are used to explain the technical solution of the present invention, and those skilled in the art can implement it based on the described technical solution and according to the resource organization method and deployment conditions of the satellite network.

[0019] This embodiment focuses on multi-demand access recommendation in satellite networks. It takes the terminal's historical access resource interaction sequence, resource identifier, and resource semantic attributes as input, and outputs the access recommendation result for the next time step through dual-branch demand representation extraction, dual-channel relationship graph construction and propagation, and candidate access resource matching.

[0020] Access resources can be satellites, beams, gateways, satellite network service resources, or resource units formed by a combination of satellite identifiers, beam identifiers, and service attributes. For access resources of different granularities, a resource identifier is set to uniquely identify the resource, and a correspondence is established between the resource identifier and resource attributes. For example, when a beam service unit is used as an access resource, the resource identifier is used to distinguish different beam service units, and the resource attributes are used to describe its satellite, coverage area, service type, bandwidth, latency, reliability, and service capabilities.

[0021] In this embodiment, the communication requirement representation is a potential requirement vector extracted from the terminal's historical access behavior and the attributes of the resources involved. Different requirement components are used to represent different access preferences or service requirements in historical interactions, and it is not required to pre-name each requirement component as bandwidth requirement, latency requirement, or reliability requirement. The specific representation of the requirement components is obtained through model training and maintains a consistent index order in the subsequent relationship graph construction and resource matching process.

[0022] Figure 1 The steps of this method are shown; Figure 2 The overall process of historical interactive input, dual-branch encoding, dual-channel relationship graph propagation, candidate resource selection, recommendation ranking and model training is shown. Figure 3 The composition and data transfer relationships of the semantic-aware relationship graph recommendation model are shown. Figure 4 A matrix representation of the differences between collaborative, competitive, and dual-channel relationships is shown.

[0023] During deployment, the terminal access control module is responsible for generating historical access records and receiving recommendation results, while the satellite network control plane is responsible for providing status information such as resource visibility, beam coverage, and basic service capabilities. The online recommendation model can be deployed on the terminal side, edge computing nodes, or network-side recommendation servers.

[0024] Historical access records can be temporarily stored in the terminal cache or written to the service-side time-series storage according to the terminal identifier. When records are uploaded, the timestamp, resource identifier, and record sequence number are retained, ensuring that historical interactions are organized in a consistent order during the training and online service phases. The trained model parameters and resource attribute encoding table are loaded by the online recommendation model, and the online service performs recommendation calculations based on currently available historical interactions and resource status.

[0025] This embodiment includes the following steps: S1. Obtain the historical access resource interaction sequence of the terminal sorted by time, integrate the identifier embedding and semantic attribute embedding of each access resource, and superimpose the position embedding to obtain the sequence representation; S2. Perform time-series coding and multi-demand routing on the sequence representation to obtain the communication demand representation; S3. Calculate the collaborative similarity between the communication requirement representations, extract the semantic requirement representation using the semantic attribute embedding and calculate the semantic similarity, and weightedly fuse the two similarities to form a hybrid similarity matrix; S4. Using the communication requirements as nodes, construct a collaborative relationship graph and a competitive relationship graph according to the high and low similarity in the hybrid similarity matrix, respectively. S5. Based on the cooperative relationship diagram and the competitive relationship diagram, the communication requirement representation is propagated to obtain the cooperative representation and the competitive representation, and then fused with the residual of the communication requirement representation to obtain the refined requirement representation; S6. Filter candidate access resources that meet the conditions of terminal visibility, beam coverage and basic service capabilities, calculate the matching score between each candidate access resource and the refined requirement representation, obtain the recommendation score through requirement attention aggregation, sort according to the recommendation score, and output the access recommendation result for the next time step.

[0026] For S1, specifically, the terminal identifier, interaction timestamp, actual access resource identifier, and corresponding resource attributes are obtained and arranged in ascending order of timestamps to form a historical access resource interaction sequence. The actual access resource can be recorded by the access control module after successful access, or it can be read from the historical access log stored in the network control plane.

[0027] For the same terminal, select the segment closest to the current recommendation time. The first valid interaction is used as the model input. Let the historical interaction sequence be:

[0028] in, Indicates the terminal. Indicates the first The access resource identifier corresponding to each interaction location. This indicates the maximum length of the input sequence. Different positions in the sequence can correspond to the same resource, and different valid interactions formed by repeatedly accessing the same resource are still retained to reflect the changes in access behavior over time.

[0029] During preprocessing, the validity of the timestamp is first verified, followed by the correspondence between the resource identifier and the attribute record in the resource directory. When multiple records exist under the same timestamp, they are arranged stably according to their sequence number or arrival order. For duplicate uploads of the same interaction record, deduplication is performed based on the record identifier to prevent the same event from being counted repeatedly in the historical sequence.

[0030] For records where the business type field is empty but the resource identifier and resource attribute can be effectively bound, the record can be retained and a missing flag can be set for the business type field; for records where the resource identifier and resource attribute cannot be established, they should be treated as unusable records to avoid merging the embedding of one resource identifier with the semantic attribute embedding of another resource.

[0031] When the number of valid interactions is insufficient When padding is used to maintain the length of the input tensor, a padding mask is set. Padding positions are only used to unify the input dimension and do not participate in attention calculation or requirement aggregation as actual interactions. When the number of effective interactions exceeds... At that time, keep the most recent one. One effective interaction.

[0032] The access resources are satellites, beams, gateways, satellite network service resources, or resource units formed by a combination of satellite identifiers, beam identifiers, and service attributes; The semantic attribute embedding is obtained by encoding at least one of the following attributes: resource type, service description, business category, bandwidth, latency, reliability, coverage area, and service capability. Continuous attributes are discretized, labeled, or jointly encoded with text attributes before participating in the encoding. The identifier embedding and semantic attribute embedding are achieved by mapping the semantic attribute embedding to the identifier embedding space through a semantic projection matrix, and then multiplying the projection result by a semantic fusion coefficient and adding it to the identifier embedding.

[0033] Specifically, the corresponding vector is read from the identifier embedding table according to the resource identifier, and a semantic attribute embedding is generated based on the semantic attributes of the resource. Discrete attributes such as resource type, business category, and service capability category can be converted into embedding vectors through a preset attribute encoding table; continuous attributes such as bandwidth, latency, and reliability can be discretized according to a predetermined interval, or they can be converted into corresponding attribute labels and then participate in semantic encoding.

[0034] For text attributes such as service descriptions, a semantic representation can be formed first, and then combined with the encoded results of other structured attributes. The encoded results of various attributes are organized according to a fixed field order, so that the same attribute meaning and input position are used in the training phase and the online service phase.

[0035] Resource The identifier is embedded as Semantic attribute embedding as Then resource integration can be represented as:

[0036] in, This is the semantic projection matrix, used to map semantic attribute embeddings to the vector space where the identifier embeddings reside; For semantic fusion coefficients, used to adjust the contribution of semantic attribute information to resource representation; This represents the merged resource.

[0037] Identifier embedding is used to distinguish the identities of different resources, while semantic attribute embedding is used to express the service capabilities and application attributes of a resource. Through this fusion, the same resource representation simultaneously contains resource identity information and resource semantic information.

[0038] The resource fusion representation is written to its corresponding historical interaction position, and then the position embedding is superimposed to obtain:

[0039] in, For the first A sequence representation of the interaction positions. This represents the fused representation of the resource corresponding to this location. Position embedding is used. The representation dimension of each position is unified as follows: This forms the sequence representation matrix of the input timing encoder.

[0040] The resource attribute encoding table is established based on the resource catalog during the training phase and loaded with the same version during the online phase. The discrete ranges of continuous attributes, the meaning of attribute labels, and the text encoding methods remain consistent between the training and online phases to avoid inconsistent representations of the same resource attributes in the two phases.

[0041] For newly emerging resources with complete attributes, semantic encoding is first completed according to the defined fields, and resource indexing and identifier embedding are synchronized. Only when the representations required for model calculation have been established and the dimensions are consistent will the resource be included in the resource catalog that can participate in the calculation.

[0042] For S2, specifically, the sequence representation formed by S1 is input into the timing encoder to encode the chronological relationship of historical interactions. Then, the encoding results of each interaction position are distributed to different demand components through multi-demand routing to obtain the communication demand representation.

[0043] The processing of the sequence representation includes: The sequence representation is encoded using a Transformer encoder with causal masking, so that each interaction position uses only the information of the current interaction position and the previous positions; The encoding results are softly assigned to a preset number of demand components through a routing mechanism, and then aggregated to obtain the communication demand representation.

[0044] Specifically, the Transformer encoder calculates attention based on the interaction position and masks interactions after the current position using a causal mask. For the ... The encoding result of each position depends only on the first to the last position. This allows for the use of information at each location, ensuring that the range of information used during the training phase is consistent with the range of information available during online recommendation.

[0045] Causal masks and padding masks are used together. Causal masks are used to exclude future positions, while padding masks are used to exclude padding positions that do not correspond to actual interactions. Before attention normalization, invisible positions are masked so that they do not participate in the information aggregation between effective interactions.

[0046] The encoding results of valid locations are input into the multi-demand routing module. The routing module sets a preset number of demand components, calculates the allocation weight of each coded location to each demand component, and normalizes it along the demand component dimension. An interaction location can contribute to multiple demand components simultaneously, and it is not required to be rigidly classified into a single demand category.

[0047] For the same demand component, the encoding results are weighted and aggregated according to the assigned weights at each position to form the communication demand vector corresponding to that demand component. Let the number of demand components be... The resulting communication demand representation matrix is ​​as follows:

[0048] in, Indicates by A communication demand representation matrix composed of communication demand components. Indicates the first Each communication requirement component Each line corresponds to a requirement node, with a superscript. This indicates the transpose operation. Indicates the first The transpose of each communication demand component is used as a matrix. of a line.

[0049] The routing center and routing parameters are updated using historical interaction samples during the training phase and loaded according to the published model version during the online phase. The number of demand components, index order, and representation dimensions remain consistent within the same model version, enabling subsequent similarity calculations and graph propagation to be processed according to the same demand node index.

[0050] In streaming services, newly arriving valid interactions are added to the historical cache, and the model re-extracts demand representations based on the updated historical sequence. In batch training, the same causal mask and padding mask rules are applied to each sample sequence. Each demand component retains its corresponding validity label, and demand components without valid input support do not participate in subsequent node pairing and neighbor selection.

[0051] The timing encoder also retains the encoding state corresponding to the most recent valid interaction. This state is used to represent the short-term context formed by the terminal's recent access behavior, and it participates in the candidate resource scoring in S6 together with the multi-demand aggregation result.

[0052] For S3, specifically, collaborative similarity between communication requirement representations and semantic similarity between semantic requirement representations are constructed separately, and then the two similarities are fused according to the same requirement node pair.

[0053] The communication branch reads the communication requirement representation formed by S2. For each pair of different demand components, collaborative similarity is calculated. The semantic branch reads the semantic attribute embeddings of each resource according to the historical interaction order, forming a pure semantic sequence, and obtains the semantic demand representation through semantic sequence encoding and routing.

[0054] The resource order in the pure semantic sequence is consistent with the historical interaction order of S1. Semantic branches use resource semantic attribute embeddings, without embedding resource identifiers and merging them again, in order to preserve the expressive role of the branch in resource semantic association.

[0055] The collaborative similarity is the cosine similarity between two communication requirement representations; The semantic requirement representation is extracted through a pure semantic sequence, which is taken from the semantic attribute embedding and processed by Transformer encoding and routing. The semantic similarity is the cosine similarity between two semantic requirement representations. Each element in the hybrid similarity matrix is ​​obtained by weighted sum of the collaborative similarity and semantic similarity of the same demand node pair, with the weights of the two similarities being the hybrid similarity coefficient and one minus the hybrid similarity coefficient, respectively.

[0056] Specifically, the semantic branch output corresponds to the communication branch. Each semantic requirement component is denoted as... And form a semantic requirement representation matrix. The two branches use corresponding demand component indices and matrix row and column order, and use the same index when calculating the similarity of the same pair of nodes. and .

[0057] The cosine similarity between two non-zero vectors is calculated using the following formula:

[0058] in, Represents the non-zero vectors involved in the cosine similarity calculation, with superscript indicating the vector. This indicates the transpose operation. Let represent the 2-norm. Therefore, the ... The demand node and the first The mixed similarity among the demand nodes is:

[0059] in, and They represent the first The first communication requirement component and the first Each communication requirement component and They respectively represent the first The first communication demand component and the first The first communication requirement component corresponds to the first The semantic requirement component and the first Each semantic requirement component and Indicates the index of the demand node; For the mixed similarity matrix The Middle line, number Column elements; For the mixed similarity coefficient, satisfying ,when When the similarity is close to 1, the hybrid similarity emphasizes the association of communication needs formed by historical access behavior; when When the value is close to 0, it places greater emphasis on the semantic association formed by resource attributes. This coefficient is related to the semantic fusion coefficient in S1. They are applied to the relational computation phase and the resource representation phase, respectively.

[0060] Before calculating similarity, the validity of the communication requirement representation and the semantic requirement representation is checked. If a requirement component is a zero vector or its corresponding position is marked as invalid, the node pair is not sent to the current neighbor sorting process to avoid invalid denominators in the cosine similarity calculation.

[0061] The communication branch and the semantic branch use the same range of valid nodes. Only when the same node is valid for both required representations will the corresponding mixed similarity be formed according to the weighted relationship described above.

[0062] The off-diagonal elements of the hybrid similarity matrix are used to describe the associations between nodes with different needs, while the diagonal elements are not considered as their own neighbors. Figure 2The collaborative similarity calculation, semantic similarity calculation, and weighted fusion in this step correspond to the calculation of the two types of relationships and the formation of the hybrid similarity, respectively.

[0063] For S4, specifically, each demand component in the communication demand representation is used as a graph node. Based on the hybrid similarity matrix, high-similarity neighbors and low-similarity neighbors are selected for each node to construct a collaborative relationship graph and a competitive relationship graph, respectively.

[0064] Collaboration diagrams are used to express the information relationships between relatively similar demand components; competition diagrams are used to express the distinguishing relationships between relatively dissimilar demand components. The competition relationship here is a representational relationship formed by the similarity between demand representations; it cannot be inferred solely from this relationship that the corresponding resources will inevitably interfere with each other or experience capacity contention at the physical layer.

[0065] In the collaborative relationship graph, a preset number of neighbors with the highest mixed similarity are retained for each demand node. The edge weight is obtained by dividing the mixed similarity by the positive temperature parameter, taking the exponent, and normalizing it among the retained neighbors. In the competitive relationship graph, a preset number of neighbors with the lowest mixed similarity are retained for each demand node. The edge weight is obtained by dividing the negative number of the mixed similarity by the positive temperature parameter, taking the exponent, and normalizing it among the retained neighbors. The edge weight corresponding to the unretained neighbors is zero.

[0066] Specifically, for demand nodes First, exclude the node itself and invalid nodes, then sort the remaining nodes from high to low according to their mixed similarity. When the similarity is the same, sort them in ascending order according to the index of the required node to ensure that the same input and the same model parameters produce a definite sorting result.

[0067] Select a predetermined number of neighbors from the beginning of the sorted list to form a set of highly similar neighbors; select a predetermined number of neighbors from the end of the list to form a set of low-similarity neighbors. Let the predetermined number of neighbors in each group be . The actual number selected is The smaller of the number of available neighbors.

[0068] When the number of available neighbors is not less than When the number of available neighbors is less than 1, the two groups of neighbors do not overlap; when the number of available neighbors is less than 1, the two groups of neighbors do not overlap. In this case, the two groups can contain the same nodes, but they calculate edge weights separately and use different propagation parameters subsequently. When there are no valid neighbors, the corresponding row in the adjacency matrix remains 0.

[0069] Set nodes The set of highly similar neighbors is Then the edge weights of the collaborative relationship graph are:

[0070] Set nodes The low-similarity neighbor set is Then the edge weights of the competition graph are:

[0071] in, It is a positive temperature parameter, which satisfies .

[0072] For cooperative graphs, a larger mixed similarity corresponds to a larger unnormalized edge weight; for competitive graphs, a smaller mixed similarity corresponds to a larger unnormalized edge weight. Normalization is performed on both sets of neighbors, using the sum of edge weights within their respective sets as the denominator.

[0073] The adjacent edge weights of the competitive relationship graph are obtained by exponential normalization and are themselves non-negative; the suppression effect of the competitive channel is achieved through residual subtraction in S5.

[0074] In numerical implementation, the maximum value within each set can be subtracted from the exponent input of each set before performing exponent calculation and normalization to reduce the possibility of numerical overflow. This process does not change the normalization result.

[0075] The adjacency matrix is ​​stored in a fixed row and column order according to the required node index. For the effective neighbor set, only the non-zero edge weights corresponding to the selected neighbors are retained, thereby transforming the mixed similarity relationship into two sets of sparse relationship graphs. The number of neighbors and temperature parameters are saved with the model configuration and remain consistent within the same online model version.

[0076] For S5, specifically, the collaborative relationship graph and the competitive relationship graph propagate the communication demand representation respectively. Both channels use the same original communication demand representation as node features, aggregate neighbor information based on their respective adjacency matrices, and transform them using non-shared parameter matrices.

[0077] The formation of the cooperative and competitive representations includes: The required node information is aggregated based on the adjacency matrix of the collaborative relationship graph, and the collaborative representation is obtained through the positive parameter matrix and ReLU activation. The demand node information is aggregated based on the adjacency matrix of the competition relationship graph, and the competition representation is obtained through the negative parameter matrix and ReLU activation. The positive parameter matrix and the negative parameter matrix are not shared, forming the cooperative representation and the competition representation.

[0078] The propagation calculation for the cooperative channel is as follows:

[0079] The propagation calculation for the contention channel is as follows:

[0080] in, and These are the adjacency matrices for the collaborative relationship graph and the competitive relationship graph, respectively. and These are the parameter matrices for the two channels, respectively. For collaborative representation, This indicates competition.

[0081] ReLU activation is performed element-wise. .

[0082] Specifically, for a given demand node, first read the non-zero edges in the adjacency matrix row where the node is located, then sum the representations of adjacent demand nodes according to their edge weights to obtain the neighbor aggregation result. Multiply the aggregation result by the parameter matrix of the corresponding channel and apply ReLU activation to form the propagation representation of the node in that channel.

[0083] The two sets of parameter matrices are maintained separately and updated according to their respective computational paths during training, and loaded separately during online propagation. The parameters of the collaborative channel and the competing channel are not interchanged, thus enabling the two channels to generate different feature transformations.

[0084] Before propagation, the dimensions of the adjacency matrix, demand representation matrix, and parameter matrix are checked. The number of rows and columns in the adjacency matrix corresponds to the number of demand nodes, and the input dimensions of the parameter matrix correspond to the feature dimensions of the demand representation. To perform subsequent residual fusion, the representation dimensions of the outputs of the two channels are kept consistent with the original communication demand representation.

[0085] The propagation results of both channels are written back according to the original required node index, and the node order is not changed due to differences in the number of neighbors or edge weights. For nodes with an empty neighbor set, the propagation result of their corresponding channel is taken as a zero vector.

[0086] The formation of the refinement requirement representation includes: The communication requirements are represented as the residual baseline; The refined requirement representation is obtained by multiplying the collaborative representation by the positive enhancement coefficient and adding it to the residual benchmark, and then subtracting the product of the competition representation and the negative suppression coefficient.

[0087] Refinement requirements are calculated using the following formula:

[0088] in, To refine the representation of requirements, It is the positive enhancement coefficient. This is the negative inhibition coefficient.

[0089] Specifically, for each demand component, its original communication demand vector is used as the residual reference, and the propagation result of the cooperative channel is multiplied by... Then add to this benchmark, and subtract the result of the competing channel propagation. The product. This operation is performed element-wise on the same demand component and the same feature location.

[0090] In one implementation, and Non-negative coefficients are used to maintain the operational direction of collaborative enhancement and competitive suppression. The specific values ​​of the coefficients are determined based on the training and validation results, and are saved and loaded together with the model parameters.

[0091] When a channel has no effective neighbors, its residual contribution to the corresponding demand component is 0, and the original communication demand representation is still retained through the residual path. When a fusion coefficient is set to 0, the direct contribution of the corresponding channel to the current residual output is 0; whether its parameters can be updated by gradient depends on whether the computational path related to that channel is still retained in the training loss.

[0092] After residual fusion, the refined requirement representation is checked to see if it has valid values ​​and correct dimensions. Requirement components that pass the check are entered into candidate resource matching, while invalid requirement components are excluded from subsequent attention normalization according to their validity flags.

[0093] For S6, specifically, based on the satellite ephemeris or network control plane information available at the current recommended time, resources that meet the conditions of terminal visibility, beam coverage and basic service capabilities are selected to form a candidate access resource set.

[0094] The candidate access resources are filtered based on the visibility, coverage, and service capability information provided by the satellite ephemeris or network control plane at the current recommended time, excluding access resources that do not meet any of the conditions for terminal visibility, beam coverage, and basic service capability.

[0095] Specifically, satellite ephemeris is used to provide information on satellite positions and their changes over time, which, combined with the terminal's position, determines the visibility of relevant resources; beam coverage information is used to determine whether the terminal is within the service range of the corresponding beam; and basic service capability information is used to determine whether the resources can provide the corresponding access services in the current state.

[0096] The network control plane can directly provide visibility flags, coverage flags, and basic service capability flags, which it has calculated. For each resource, it is added to the candidate access resource set only if all the required flags are valid. If any condition is not met, or if the necessary state for the judgment is missing, the resource is not included in the current candidate set.

[0097] Resource status at the current recommendation time is organized by timestamp. When timestamps from different status sources are inconsistent, statuses that are still valid at the current recommendation time should be used; expired statuses are not considered as a basis for resource availability. After the candidate set is formed, the same status snapshot is used in this matching and ranking calculation.

[0098] When a resource is at the boundary of the visibility window, filtering is performed based on the time of recommendation and the effective visibility status. Resources that may enter the visibility range in the future are not added to the current candidate set in advance based on their future visibility. Resources that are outside the beam coverage or have insufficient basic service capabilities, even if they have a high historical interaction frequency, still need to pass the same candidate filtering criteria.

[0099] The candidate resource selection process determines the range of resources that can participate in the ranking. The demand representations formed by S1 to S5 are used to express the terminal's historical behavior and demand relationships, while the candidate set formed by S6 is used to constrain the current resource availability. Both work together in the subsequent matching calculations.

[0100] For candidate resources Read its resource fusion representation Calculate the matching score between the resource and each refinement requirement component:

[0101] in, For the first Individualized requirements and weighting For this demand component and candidate resources Match score, superscript This indicates the transpose operation. This represents the inner product between the refined requirement component and the candidate resource representation.

[0102] The determination of the recommendation score includes: The matching score is obtained by calculating the inner product between each refined requirement representation and the resource representation of the candidate access resource; For each matching score of the same candidate access resource, a softmax operation is performed to obtain the demand attention weight, and the matching scores are weighted and aggregated according to the demand attention weight. The encoding state corresponding to the most recent interaction is used as the short-term context representation. The inner product between the short-term context representation and the resource representation of the candidate access resource is calculated to obtain the context score. The weighted aggregation result is added to the context score, and the resource bias of the candidate access resource is added to obtain the recommendation score.

[0103] Specifically, for the same candidate resource, its matching score with all valid refined demand components is collected, and softmax normalization is performed on the demand component dimension. Let the set of valid demand component indices be... ,but:

[0104] in, For the first Each demand component is for candidate resources Attention weights Indicates the first Individual Refinement Requirements and Candidate Resources The matching score between them Indicates the first Individual Refinement Requirements and Candidate Resources The matching score between them For the effective demand component index set The summation index in the table.

[0105] For different candidate resources, their demand attention weights are calculated separately. The same demand component can have different weights for different candidate resources, allowing the score of a candidate resource to be adjusted according to the degree of matching between the resource and each demand component.

[0106] Aggregate the matching scores according to the attention weights of the needs to obtain:

[0107] in, Candidate resources Multi-demand aggregated score.

[0108] Read the encoded state corresponding to the most recent valid interaction in S2 as a short-term context representation. Taking the inner product of this with the candidate resource representation, we get:

[0109] in, For short-term context scores, superscript This indicates the transpose operation. Short-term context representation With candidate resource representation The inner product between them. The dimension of the short-term context representation is consistent with the dimension of the candidate resource representation, enabling direct inner product operations between the two.

[0110] The multi-demand aggregate score and short-term context score are merged and overlaid with resource bias:

[0111] in, and These are the fusion weights for the multi-demand aggregation score and the short-term context score, respectively. Candidate resources The resource bias is such that all variables are scalars.

[0112] The multi-demand aggregation score reflects the overall matching between candidate resources and multiple demand components, while the short-term context score reflects the association between the most recent interaction state and the candidate resource. Resource bias is maintained along with the model parameters and is used in the scoring calculation of different resources.

[0113] Candidate access resources are ranked from highest to lowest according to their recommendation scores to obtain the access recommendation results for the next time step. When recommendation scores are the same, they are arranged according to the stable order of resource identifiers to maintain the determinism of the output. When multiple candidate resources need to be returned, the first few items in the sorted results are truncated according to the configured number.

[0114] The recommendation result should include at least the resource identifier, recommendation score, and timestamp of this recommendation. After receiving the recommendation result, the access control module will perform subsequent access selection based on the network status at the actual access time.

[0115] If the candidate resource set is empty, a "no recommendations to execute" status is returned, along with the corresponding timestamp and the reason for the selection. This status is not included in the supervision sample as a real access resource. If the resource status changes before actual access, the relevant resources should be re-examined based on the updated valid status; if a new recommendation needs to be formed, candidate selection and scoring should be re-executed.

[0116] like Figure 3 As shown, the embedding layer receives a sequence of resources organized according to the interaction order. Resource 1, Resource 2, Resource 3, and Resource 4 in the diagram represent resource inputs at different interaction positions. The above numbering is only for illustrating the sequence structure and does not limit the historical sequence to only 4 interactions, nor does it limit the 4 positions to corresponding to different resources.

[0117] The multi-interest extraction layer corresponds to the multi-demand extraction process in this embodiment, organizing the temporal coding results into multiple demand components through a routing mechanism. In the figure, "interest" represents potential preferences extracted from historical interactions, corresponding to communication demand representation in the satellite network scenario of this embodiment.

[0118] Figure 3 Collaborative filtering representation in The semantic representation corresponds to the communication demand representation extracted from the terminal's historical access behavior. The superscript corresponds to the semantic requirement representation extracted from the sequence of resource semantic attributes. This is used to represent the corresponding terminals. Both are first processed through a dual-channel similarity graph, and then the mixed similarity results are used to form a collaborative relationship graph and a competitive relationship graph, respectively.

[0119] The collaborative graph convolutional network and the competitive graph convolutional network propagate the communication requirement representation based on their respective adjacency matrices, forming two propagation paths using independent parameters. The two paths are then residually fused using the weighted enhancement, weighted suppression, and original representation preservation relationships of S5 to obtain the refined requirement representation. .

[0120] Figure 3 The fusion symbol in the code is used to represent the merging of the two propagation results. The specific computational relationship is performed according to the residual fusion formula in this embodiment. The cooperative channel corresponds to the positive enhancement term, and the competing channel corresponds to the negative suppression term. The original communication requirements are represented as the residual baseline and retained.

[0121] During the online service phase, Figure 3 The nearest neighbor in the S6 algorithm is used to illustrate the service steps for retrieving matching resources based on demand. The retrieval scope is limited to candidate access resources filtered by S6. The aggregation module aggregates multiple demand matching scores for the same candidate resource, and the final ranking is performed according to the recommended score calculation method of S6.

[0122] Figure 3 The training and online service phases use the same parameters as the training model. During training, parameters are updated using a loss function, and during the online phase, the trained parameters are loaded to complete representation extraction, relation propagation, and resource scoring.

[0123] like Figure 4 As shown, the demand relationship matrix diagram includes... Figure 4 (a) shows the collaborative relationship matrix, Figure 4 (b) shows the competition relationship matrix and Figure 4 (c) shows the dual-channel relationship difference matrix. The horizontal and vertical axes of each matrix are demand indices, with demand 1 to demand 5 representing example demand nodes.

[0124] The collaboration matrix displays the strength of relationships within collaborative channels, while the competition matrix displays the strength of relationships within competitive channels. Both matrices use the same set of nodes and row / column order to compare the relationships between pairs of nodes with the same demand across the two channels.

[0125] The dual-channel relationship difference matrix is ​​used to show the difference in relative strength between two types of relationships. The low, high, negative, and positive labels in the diagram are used to help understand the strength of the relationship or the direction of the difference. The specific edge weights involved in the propagation are still based on the adjacency matrix formed by S4.

[0126] Figure 4This is a schematic representation of relationship construction. The number of nodes, matrix display method, and color intensity do not constitute a limitation on the actual required number or the range of edge weight values. In actual operation, the size of the matrix is ​​determined by the number of valid required components; invalid nodes do not participate in neighbor sorting, and the edge weight corresponding to unselected neighbors is 0.

[0127] For model training, the parameters for calculating the recommendation score are obtained by training by minimizing the joint loss. The joint loss is formed by weighted summation of the recommendation loss with the real next-interaction access resource as the supervision signal, the demand diversity loss to reduce redundancy between different communication demand representations, and the cooperative-semantic alignment loss to improve the consistency between access behavior representation and resource semantic representation.

[0128] Specifically, the historical access records of terminals are organized chronologically. In the training samples, historical interactions prior to a certain moment are used as input, and the next actual access to resources after that input is used as the supervision target. The interaction corresponding to the supervision target is not included in the historical input sequence of the sample, so as to maintain the temporal relationship between the prediction object and the known input.

[0129] The training samples include historical access resource interaction sequences, corresponding resource semantic attributes, the actual next access resource, and the candidate resource status corresponding to the sample time. When training involves availability filtering, the historical state at the sample time is used to avoid using the resource state at the training runtime to replace historical availability conditions.

[0130] For each training sample, the demand representation, dual-channel propagation results, and resource recommendation score are calculated according to S1 to S6. The recommendation loss is calculated using the actual next-interaction access resource as a supervision signal, enabling the model to learn the correspondence between historical interactions and the next access resource.

[0131] Demand diversity loss applies to different communication demand components, constraining the redundancy in representation between components. This constraint reduces the likelihood of multiple demand components expressing the same historical information, enabling multi-demand routing outputs to retain distinct demand information.

[0132] The collaborative-semantic alignment loss is applied to the access behavior representation and the resource semantic representation. It establishes an alignment relationship through the two types of representations of corresponding resources and forms a distinguishing constraint by combining non-corresponding representations, so that behavioral information and resource attribute information can play a role together in model training.

[0133] The combined loss is ,in, To recommend losses, Due to the loss of demand diversity, For collaborative-semantic alignment loss; and The corresponding loss weights are used to adjust the contribution of each auxiliary constraint to the model update.

[0134] The above loss weights and mixed similarity coefficients They are used separately, without confusing their calculation positions and functions. The hybrid similarity coefficient is used to fuse the similarity of the two types of nodes, and the loss weights are used to combine the training objectives.

[0135] In one specific training implementation, the recommendation model is built based on the PyTorch deep learning framework. The model sequentially includes a resource representation input part, a communication demand extraction branch, a semantic demand extraction branch, a dual-channel relationship graph propagation part, and a recommendation scoring part. The communication demand extraction branch uses a two-layer Transformer encoder with causal masks to temporally encode the historical access resource interaction sequence and forms multiple communication demand components through dynamic routing. The semantic demand extraction branch performs Transformer encoding and routing processing on the pure semantic sequence formed by embedding resource semantic attributes to obtain semantic demand components corresponding to the communication demand components according to their demand node indices. The outputs of the two branches are used to form a hybrid similarity matrix and to construct a collaborative relationship graph and a competitive relationship graph, respectively. The two relationship graphs use independent graph propagation parameters to process the communication demand representation. The resulting collaborative and competitive representations are residually fused with the original communication demand representation to form a refined demand representation, which is then matched with the resource representation of the candidate access resources and combined with short-term context to obtain a recommendation score. In this implementation, the maximum length of the historical access resource interaction sequence is set to 50, the resource representation dimension is set to 64, the number of demand components K is set to 5, and each effective demand node selects 2 high-similarity neighbors and 2 low-similarity neighbors. The semantic fusion coefficient γ is set to 0.3, the mixed similarity coefficient λ is set to 0.6, and Dropout is set to 0.1. During training, the Adam optimizer is used, the batch size is set to 128, the number of training rounds is set to 100, and the initial learning rate is set to 0.001. After the 20th round of training, the learning rate is multiplied by 0.5 every 20 rounds, and the lower limit of the learning rate is set to 0.00001. Each training batch performs forward computation based on the historical input to obtain the recommendation loss, demand diversity loss, and collaborative-semantic alignment loss, forming the joint loss. Backpropagation and parameter updates are performed based on the joint loss, and training is completed sequentially for each batch until the preset number of training rounds is reached. For resource semantic attributes in text form, such as service descriptions, semantic embeddings can be pre-generated using the all-MiniLM-L6-v2 pre-trained text encoding model. The pre-generated semantic embeddings remain fixed during the training of the recommendation model.

[0136] During training, the gradients of relevant parameters are calculated based on the joint loss, and the trainable parameters involved in resource embedding, temporal coding, demand routing, semantic projection, graph propagation, and recommendation scoring are updated. Configuration items not set as trainable variables are used according to their preset configurations and recorded with the model version.

[0137] After training is complete, save the model parameters, resource index, resource attribute encoding table, number of requirement components, and parameter dimension information. During online loading, first check if the versions and dimensions of these contents correspond; if the check passes, load the complete model version.

[0138] If the new version's parameters or resource attribute encoding table fails to load, continue using the previous valid version. Avoid mixing the embedding of identifiers, semantic encoding, or graph propagation parameters from different versions. Model updates can be triggered according to the release cycle or changes in the resource catalog, but the online service should only be switched after the new version has been fully loaded.

[0139] The online service phase involves updating historical records, calculating demand representations, propagating relationship graphs, filtering candidate resources, and ranking recommendations. Gradients and optimizer states generated during the training phase are not considered inputs to the online recommendation system. The online recommendation outputs and subsequent real-world access results can continue to form new historical records for use in subsequent training.

[0140] Figure 2 The “Model Score and Representation” in the text correspond to the score results and intermediate representations read from the training loss; the “Model Parameters After Training” correspond to the parameters saved and loaded into the online recommendation model after joint loss optimization, thus forming the parameter update relationship between training and online service.

[0141] In one optional configuration, the input sequence is truncated from the 50 most recent valid interactions. Both the resource identifier embedding and the projected semantic attribute embedding are represented using 64 dimensions, and multi-demand routing forms 5 demand components. The hybrid similarity coefficient is set to 0.6, and each valid demand node selects 2 high-similarity neighbors and 2 low-similarity neighbors.

[0142] In this configuration, each demand node has a maximum of four other demand nodes. Therefore, when all five demand components are valid, two neighbors can be selected from the beginning and end of the sorted list, respectively. The edge weights are normalized for each of the two channels, and independent parameters are used for graph propagation and residual fusion.

[0143] The input length, representation dimension, number of required components, similarity coefficient, and number of neighbors described above are for illustrating feasible parameter configurations and do not indicate that specific experimental performance has been achieved with this configuration. In actual implementation, the appropriate configuration can be determined by combining historical interaction lengths, resource catalog size, and the computing power of the deployed devices.

[0144] In another implementation, resource semantic attributes are derived from satellite resource catalogs, beam service catalogs, or service capability tables maintained by the network control plane. As long as the relevant attributes can be mapped to the adopted semantic space according to consistent field meanings and encoding methods, the resource representation fusion, semantic similarity calculation, and candidate resource matching processes can be reused.

[0145] When the granularity of resource organization changes, the resource identifier and its attribute binding relationship are adjusted accordingly. For example, when satellites and beams are combined to form resource units, the combined identifier is used to distinguish different resource units, and its semantic attributes describe the coverage and service capabilities of the unit; when satellite network service resources are used as resource units, their semantic attributes describe the corresponding service category and capabilities, and candidate selection is performed based on the visibility, coverage, and basic service capability status associated with the service resource.

[0146] From the perspective of overall processing relationships, this method first organizes historical access behavior and resource semantic attributes into a unified sequence representation, then extracts multiple communication demand components and their corresponding semantic demand components; constructs demand relationships by fusing two types of similarity, and uses high and low similarity neighbors to form collaborative channels and competitive channels respectively.

[0147] The collaborative channel aggregates information between similar demands, while the competing channel forms a propagation representation to distinguish dissimilar demands. The two results are then refined into a demand representation through weighted enhancement, weighted suppression, and preservation of the original representation. Subsequently, this representation is matched with resources that currently meet the visibility, coverage, and basic service capability conditions, and a resource ranking is formed by combining short-term context.

[0148] Therefore, historical demand expression, handling of relationships between demands, and current resource availability are sequentially linked in the same recommendation process. The model output provides a basis for resource selection for subsequent access control, while the actual access operation is still performed based on the effective network status at that time.

[0149] The present invention has been described above with reference to specific embodiments. Based on the technical solution of the present invention, any corresponding adjustments made by those skilled in the art to the resource organization granularity, parameter configuration, and deployment location can be implemented while maintaining the above-described processing relationships.

Claims

1. A method for recommending semantic dual-channel multi-demand access in satellite networks, characterized in that, include: S1. Obtain the historical access resource interaction sequence of the terminal sorted by time, integrate the identifier embedding and semantic attribute embedding of each access resource, and superimpose the position embedding to obtain the sequence representation; S2. Perform time-series coding and multi-demand routing on the sequence representation to obtain the communication demand representation; S3. Calculate the collaborative similarity between the communication requirement representations, extract the semantic requirement representation using the semantic attribute embedding and calculate the semantic similarity, and weightedly fuse the two similarities to form a hybrid similarity matrix; S4. Using the communication requirements as nodes, construct a collaborative relationship graph and a competitive relationship graph according to the high and low similarity in the hybrid similarity matrix, respectively. S5. Based on the cooperative relationship diagram and the competitive relationship diagram, the communication requirement representation is propagated to obtain the cooperative representation and the competitive representation, and then fused with the residual of the communication requirement representation to obtain the refined requirement representation; S6. Filter candidate access resources that meet the conditions of terminal visibility, beam coverage and basic service capabilities, calculate the matching score between each candidate access resource and the refined requirement representation, obtain the recommendation score through requirement attention aggregation, sort according to the recommendation score, and output the access recommendation result for the next time step.

2. The method according to claim 1, characterized in that, The access resources are satellites, beams, gateways, satellite network service resources, or resource units formed by a combination of satellite identifiers, beam identifiers, and service attributes; The semantic attribute embedding is obtained by encoding at least one of the following attributes: resource type, service description, business category, bandwidth, latency, reliability, coverage area, and service capability. Continuous attributes are discretized, labeled, or jointly encoded with text attributes before participating in the encoding. The identifier embedding and semantic attribute embedding are achieved by mapping the semantic attribute embedding to the identifier embedding space through a semantic projection matrix, and then multiplying the projection result by a semantic fusion coefficient and adding it to the identifier embedding.

3. The method according to claim 1, characterized in that, The processing of the sequence representation includes: The sequence representation is encoded using a Transformer encoder with causal masking, so that each interaction position uses only the information of the current interaction position and the previous positions; The encoding results are softly assigned to a preset number of demand components through a routing mechanism, and then aggregated to obtain the communication demand representation.

4. The method according to claim 1, characterized in that, The collaborative similarity is the cosine similarity between two communication requirement representations; The semantic requirement representation is extracted through a pure semantic sequence, which is taken from the semantic attribute embedding and processed by Transformer encoding and routing. The semantic similarity is the cosine similarity between two semantic requirement representations. Each element in the hybrid similarity matrix is ​​obtained by weighted sum of the collaborative similarity and semantic similarity of the same demand node pair, with the weights of the two similarities being the hybrid similarity coefficient and one minus the hybrid similarity coefficient, respectively.

5. The method according to claim 1, characterized in that, In the collaborative relationship graph, a preset number of neighbors with the highest mixed similarity are retained for each demand node. The edge weight is obtained by dividing the mixed similarity by the positive temperature parameter, taking the exponent, and normalizing it among the retained neighbors. In the competitive relationship graph, a preset number of neighbors with the lowest mixed similarity are retained for each demand node. The edge weight is obtained by dividing the negative number of the mixed similarity by the positive temperature parameter, taking the exponent, and normalizing it among the retained neighbors. The edge weight corresponding to the unretained neighbors is zero.

6. The method according to claim 1, characterized in that, The formation of the cooperative and competitive representations includes: The required node information is aggregated based on the adjacency matrix of the collaborative relationship graph, and the collaborative representation is obtained through the positive parameter matrix and ReLU activation. The demand node information is aggregated based on the adjacency matrix of the competition relationship graph, and the competition representation is obtained through the negative parameter matrix and ReLU activation. The positive parameter matrix and the negative parameter matrix are not shared, forming the cooperative representation and the competition representation.

7. The method according to claim 1, characterized in that, The formation of the refinement requirement representation includes: The communication requirements are represented as the residual baseline; The refined requirement representation is obtained by multiplying the collaborative representation by the positive enhancement coefficient and adding it to the residual benchmark, and then subtracting the product of the competition representation and the negative suppression coefficient.

8. The method according to claim 1, characterized in that, The candidate access resources are filtered based on the visibility, coverage, and service capability information provided by the satellite ephemeris or network control plane at the current recommended time, excluding access resources that do not meet any of the conditions for terminal visibility, beam coverage, and basic service capability.

9. The method according to claim 1, characterized in that, The determination of the recommendation score includes: The matching score is obtained by calculating the inner product between each refined requirement representation and the resource representation of the candidate access resource; For each matching score of the same candidate access resource, a softmax operation is performed to obtain the demand attention weight, and the matching scores are weighted and aggregated according to the demand attention weight. The encoding state corresponding to the most recent interaction is used as the short-term context representation. The inner product between the short-term context representation and the resource representation of the candidate access resource is calculated to obtain the context score. The weighted aggregation result is added to the context score, and the resource bias of the candidate access resource is added to obtain the recommendation score.

10. The method according to claim 1, characterized in that, The parameters for calculating the recommendation score are obtained by training by minimizing the joint loss. The joint loss is formed by weighted summation of the recommendation loss with real next-interaction access resources as supervision signals, the demand diversity loss that reduces redundancy between different communication demand representations, and the cooperative-semantic alignment loss that improves the consistency between access behavior representation and resource semantic representation.

Images