A resource regulation method for semantic perception guidance and structure denoising in a computing power network

Abstract

The application discloses a resource regulation method for semantic perception guidance and structure denoising in a computing power network, comprising the following steps: obtaining perception data of a task request and computing power resources, performing subgraph sampling of topology maintenance to construct a perception correlation subgraph; constructing an asymmetric twin network containing a target branch and an online branch based on the perception correlation subgraph; applying a weak disturbance strategy to the perception correlation subgraph to construct a target view, and applying a strong disturbance strategy to construct a defective view; inputting the target view into the target branch to generate a semantic perception matching matrix, inputting the defective view into the online branch, and calculating a cross-view resource matching prediction matrix by obtaining demand adaptation characteristics of the online branch and the target branch; constructing a reconstruction loss function based on the semantic perception matching matrix and the resource matching prediction matrix, updating the online branch parameters based on the reconstruction loss function, and applying the converged online branch to real-time collected computing power network perception data to output a regulation matching relationship between the task request and the computing power resources.

Description

Technical Field

[0001] This invention belongs to the field of computing power network perception and resource management technology, and particularly relates to a resource regulation method for semantic perception guidance and structural denoising in computing power networks. Background Technology

[0002] Resource scheduling algorithms based on Graph Neural Networks (GNNs) are widely used in processing graph data relating tasks and resources. These methods typically model task requests and computing resources as nodes in a graph, and the mapping between task allocation and resource scheduling as edges, learning resource adaptation preferences by aggregating neighbor information. To improve the ability to perceive large-scale computing network topologies, some existing technologies introduce clustering mechanisms to divide resource states into groups, using K-means or Sinkhorn algorithms to map nodes with similar features to the same semantic prototype to obtain richer group-perceived semantic information. Furthermore, graph contrastive learning, as a mainstream paradigm of self-supervised learning, constructs different enhanced views such as feature masks or edge perturbations, and uses negative sampling strategies to maximize the similarity of positive sample pairs and widen the distance between negative sample pairs, thereby learning robust task and resource representations under unlabeled conditions. However, the above methods still face significant challenges when dealing with complex scenarios prevalent in real-world computing networks, such as perceptual noise, ultra-large-scale topologies, and implicit association misjudgments.

[0003] Existing technologies still have the following significant drawbacks in practical applications: First, insufficient noise resistance and overfitting to perceived noise. Traditional scheduling models directly use raw 0 and 1 hard-sensing records as supervision signals. However, in real-world computing networks, instantaneous load fluctuations of nodes and network link jitter often cause spurious peak noise or measurement errors. Forcing the model to fit such noisy data will lead to the model learning incorrect mapping patterns, making it difficult to capture the true potential correlation between tasks and resources. Second, computational and memory bottlenecks in large-scale topologies. Although clustering algorithms such as Sinkhorn can extract semantic information, they usually require global alignment and double random normalization of the entire graph's node representation in each iteration. As the number of edge nodes increases dramatically, computational complexity and memory overhead grow non-linearly, easily leading to memory overflow and computational waste. Third, negative sampling dependency and path misjudgment risks. Graph contrastive learning relies heavily on negative sampling strategies to maintain representation uniformity. However, in the context of implicit association in computing networks, randomly selected negative samples are very likely to be the potential optimal matching paths within the system. This operation of forcibly distancing semantically similar node representations can lead to serious semantic conflicts. Furthermore, searching for high-quality negative samples among a massive number of nodes can generate high computational overhead, making it difficult to meet the requirements of real-time control.

[0004] Therefore, this invention proposes a resource regulation method for semantic awareness guidance and structural denoising in computing power networks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a resource regulation method for semantic awareness guidance and structural denoising in computing power networks, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a resource regulation method for semantic awareness guidance and structural denoising in computing power networks, comprising: Acquire the perception data of task requests and computing resources, and perform topology-preserving subgraph sampling to construct a perception-related subgraph; An asymmetric twin network containing target branches and online branches is constructed based on the perceptual association subgraph; A weak perturbation strategy is applied to the perceptual association subgraph to construct a target view, and a strong perturbation strategy is applied to construct a fragmented view; The target view is input into the target branch to generate a semantic-aware matching matrix. The incomplete view is input into the online branch to extract encoded features. The cross-view resource matching prediction matrix is ​​calculated by obtaining the demand adaptation features of the online branch and the target branch. A reconstruction loss function is constructed based on the semantic perception matching matrix and the resource matching prediction matrix. The online branch parameters are updated based on the reconstruction loss function, and the converged online branch is applied to the real-time collected computing power network perception data to output the control and matching relationship between task requests and computing power resources.

[0007] Optionally, the process of constructing a perceptual association subgraph includes: Based on the acquired task request and computing resource perception data, an original perception association graph is constructed, and an enhanced adjacency matrix is ​​obtained by adding self-loops to the adjacency matrix of the original perception association graph. After feature normalization of the enhanced adjacency matrix, a graph partitioning algorithm is used to partition the entire set of nodes of the original perceptual association graph with the feature normalization to obtain several clusters. The set of nodes in the subgraph is determined based on a subset of clusters randomly selected in each training iteration; A perceptual association subgraph is obtained based on the set of subgraph nodes.

[0008] Optionally, the weak perturbation strategy includes setting a first feature mask rate and a first association mapping edge deletion rate; The strong perturbation strategy includes setting a second feature mask rate and a second association mapping edge deletion rate; The second feature mask rate is greater than the first feature mask rate, and the second association mapping edge deletion rate is greater than the first association mapping edge deletion rate.

[0009] Optionally, the process of generating a semantically aware matching matrix by inputting the target view into the target branch includes: The target branch encoding features are obtained by using a computing power topology encoder for the target branch to extract features from the target view. The target branch coding features are subjected to L2 normalization to obtain normalized coding features; The Sinkhorn-Knopp algorithm is used to perform online resource state self-clustering on the normalized encoded features and the learnable resource prototype vector to obtain the probability distribution matrix. The semantic-aware matching matrix is ​​obtained by calculating the potential association probability between nodes based on the probability distribution matrix.

[0010] Optionally, the process of using the Sinkhorn-Knopp algorithm to perform online resource state self-clustering on the normalized encoded features and learnable resource prototype vectors to obtain the probability distribution matrix includes: The perceptual matching log probability matrix is ​​obtained by scaling the dot product similarity between the normalized encoded features and the learnable resource prototype vector with a temperature coefficient. The initial perception assignment matrix is ​​obtained by exponentializing the perception matching log probability matrix. The initial sensing allocation matrix is ​​subjected to alternating row normalization and column normalization operations until convergence, resulting in a probability distribution matrix.

[0011] Optionally, the process of extracting encoded features from the incomplete view into the online branch and calculating the cross-view resource matching prediction matrix by obtaining the demand adaptation features of the online branch and the target branch includes: An online branching computational topology encoder is used to extract features from the incomplete view to obtain online branching coding features; The target branch requirement adaptation features are obtained by mapping the target branch encoding features using a shared architecture resource regulation decoder. A resource regulation decoder with a shared architecture is used to map the online branch encoding features to obtain online branch demand adaptation features; The target branch demand adaptation features and online branch demand adaptation features are both subjected to L2 normalization. The transpose dot product of the normalized online branch demand adaptation features and the normalized target branch demand adaptation features is calculated and then subjected to a dimension softmax operation to obtain the resource matching prediction matrix.

[0012] Optionally, the resource regulation decoder is composed of a multi-layer sensing mechanism.

[0013] Optionally, the reconstruction loss function is: ; In the formula, To reconstruct the loss value, This is the resource matching prediction matrix, where N is the number of nodes. This is a semantic-aware matching matrix. The first The and the first Index of each node.

[0014] Optionally, the process of updating the online branch parameters based on the reconstruction loss function includes: The gradient with respect to the online branch parameters is calculated based on the reconstruction loss function; The optimized online branch parameters are updated by gradient updating the online branch parameters.

[0015] Optionally, an exponential moving average strategy is used to process the updated online branch parameters to obtain the updated target branch parameters, wherein the update expression for the target branch parameters is: ; In the formula, The momentum coefficient, For online branch parameters, For the updated target branch parameters, The parameters for the target branch before the update.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs an asymmetric twin network architecture, replacing noisy original perceptual records with a semantically aware matching matrix generated by the target branch as the supervision target. This guides online branches to reconstruct potential association states from strongly perturbed incomplete views, effectively filtering out perceptual noise introduced by load fluctuations and measurement errors in the computing network, and avoiding model overfitting and incorrect mapping patterns. Simultaneously, a topology-preserving subgraph sampling strategy is employed to reduce the dimensionality of the entire graph computation to the subgraph level, significantly reducing the computational complexity and memory overhead under large-scale computing network topologies. This method maintains training stability without negative sampling, avoiding semantic conflicts caused by mistakenly identifying potential optimal paths as negative samples, ultimately achieving high-precision and highly robust real-time matching of task requests and computing resources. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a system architecture diagram of an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] Example 1 This embodiment provides a resource regulation method for semantic awareness guidance and structural denoising in a computing power network. This method constructs an asymmetric Siamese network architecture and establishes a self-supervised learning method with feature alignment between online branches and target branches. Utilizing the structural differences between views and the asymmetric parameter update mechanism, it effectively prevents model collapse under training without negative sampling. The target branch, updated with momentum, generates a semantic awareness matching matrix reflecting the association of potential perceptual states. This matrix is ​​then used to replace the noisy original association records as the reconstruction target, guiding the online branch to learn resource adaptation representations from the incomplete structure. This achieves efficient and accurate task and resource matching scheduling without the need for negative sampling.

[0021] like Figure 1 and Figure 2 As shown, this invention provides a resource regulation method for semantic awareness guidance and structural denoising in computing power networks. This method primarily addresses the problems of instantaneous load fluctuations and sensing measurement noise commonly found in sensing data of large-scale computing power networks. First, it performs a subgraph sampling operation based on topology preservation to reduce the memory overhead of full-graph computation. Then, it constructs an asymmetric Siamese network architecture and uses the target branch updated with momentum to generate a semantic awareness matching matrix reflecting the association of potential sensing states. This matrix is ​​then used to replace the noisy original association records as the reconstruction target, guiding the online branch to learn resource adaptation representations from the incomplete structure. This achieves efficient and accurate task and resource matching scheduling without the need for negative sampling. The method includes the following steps: Step 1: Obtain the original task request and computing resource awareness data, perform topology-preserving subgraph sampling to construct the awareness association subgraph, and construct an asymmetric twin awareness network architecture containing the target branch and online branch.

[0022] 1) Definition and preprocessing of task requests and computing resource awareness data.

[0023] Define the original perceptual association graph as .in Indicates inclusion A collection of task request nodes and computing resource nodes; This represents the set of mappings between task requests and computing resources, describing the connection relationships such as task offloading and resource mapping between task requests and computing resources; This represents a feature matrix containing task requirement characteristics and computing resource attributes, where This represents the original feature dimension.

[0024] Before inputting the data into the model, the raw sensory data undergoes standardization preprocessing. First, a self-loop addition operation is performed: the raw sensory adjacency matrix... Add a self-loop, i.e. ,in This is the identity matrix. The purpose of this operation is to ensure that the graph neural network can retain the task request's own feature information while aggregating neighbor node information. Then, feature normalization is performed: [the feature matrix is ​​then normalized]. Row normalization is performed to scale the feature vector of each node to a unit length, thereby eliminating the numerical differences caused by resource nodes of different sizes and accelerating subsequent training convergence.

[0025] 2) Topology-preserving subgraph sampling. To address the issue of excessive GPU memory consumption during training with large-scale network topology data, a clustering-based subgraph sampling algorithm is used to divide the entire graph into multiple subgraphs.

[0026] Furthermore, the process of constructing the perceptual association subgraph includes: constructing an original perceptual association graph based on the acquired task request and computing resource perception data; adding self-loops to the adjacency matrix of the original perceptual association graph to obtain an enhanced adjacency matrix; after feature normalization of the enhanced adjacency matrix, using a graph partitioning algorithm to partition the full graph node set of the feature-normalized original perceptual association graph to obtain several clusters; determining the subgraph node set based on a subset of clusters randomly selected in each training iteration; and obtaining the perceptual association subgraph based on the subgraph node set.

[0027] To address the memory overflow issue faced by large-scale computing networks with widely distributed edge nodes during full-graph training, this embodiment employs a clustering-based subgraph sampling strategy. The specific implementation process is as follows: First, a graph partitioning algorithm (such as METIS algorithm or spectral clustering algorithm) is used to partition the entire graph node set. Divided into Clusters with tight internal connections and sparse inter-cluster connections implicitly represent a set of related resources with similar needs, denoted as . .

[0028] Then, in each training iteration, randomly select Each cluster constructs a perceptual association subgraph for the current batch. .in, For the selected The union of all nodes in a cluster; for Middle connection The key to mapping association sets of internal nodes lies in... It not only includes edges within these clusters but also strictly preserves perceptual mapping connections between selected clusters. This strategy reduces computational complexity from... Reduce to ( (This represents the number of nodes in the subgraph), while preserving the topological structure that adapts to local needs and resources to the greatest extent possible, providing a complete structural context for subsequent perceptual clustering tasks.

[0029] 3) Asymmetric twin view construction: Apply a weak perturbation strategy to the perceptual correlation subgraph to construct the target view, and apply a strong perturbation strategy to construct the incomplete view.

[0030] Perceptual association subgraph for sampling This embodiment constructs a branch containing the target branch (corresponding to...) Figure 2 (Top branch) and online branch (corresponding) Figure 2 The asymmetric twin network architecture (within the lower branch) is used, and differentiated perturbation strategies are applied to each branch, i.e., weak perturbation strategies are designed separately. and strong perturbation strategy To construct training samples of varying difficulty: In the target branch (target view), a weak perturbation strategy is adopted. Set a lower first feature mask rate. and a lower first association mapping edge deletion rate The aim is to preserve the original structure of the subgraph to the greatest extent possible, that is, to retain the original computational network state information of the subgraph as much as possible, so as to ensure that the generated clustering target has a high confidence level.

[0031] Regarding online branches (incomplete views): a strong perturbation strategy is employed. Set a higher second feature mask rate. Second association mapping edge deletion rate ,satisfy and The specific steps are as follows, for the feature matrix Randomly occlude some node features; for the perceptual adjacency matrix of the subgraph The model randomly discards some topological links, generating a perceptual adjacency matrix with a severely incomplete topological structure. This incomplete matrix can be used for reconstruction training, enabling online branches to infer the potential semantic links between task requests and computing resources even from the incomplete structure, thus improving the stability of resource allocation in noisy environments. This process simulates anomalies such as lost perceptual measurements and random load fluctuations in real-world scenarios, guiding the online branch model to recover potential adaptive semantics from damaged association structures.

[0032] Step 2: Construction of semantic-aware matching matrix and resource matching prediction matrix: Forward propagation is performed based on dual-view data. The semantic-aware matching matrix is ​​generated using the target branch, and the resource matching prediction matrix across views is calculated using the online branch.

[0033] 1) Generate a semantically aware matching matrix using the target branch (corresponding to the target view). The process includes: using the computing power topology encoder of the target branch to extract features from the target view to obtain the target branch encoded features; performing L2 normalization on the target branch encoded features to obtain normalized encoded features; using the Sinkhorn-Knopp algorithm to perform online resource state self-clustering on the normalized encoded features and the learnable resource prototype vector to obtain a probability distribution matrix; and calculating the potential association probability between nodes based on the probability distribution matrix to obtain the semantically aware matching matrix.

[0034] In this specific implementation, the target branch acts as a "teacher" in this invention, responsible for generating high-quality resource matching targets. This process does not calculate gradients, and its parameters... It follows the online branch evolution solely through a momentum update mechanism. The specific implementation process is as follows: Feature extraction from computing power topology encoder: Inputting the weakly perturbated target view into the computing power topology encoder of the target branch yields the encoded features of task requests and computing resources. The encoder can employ a multi-layer graph convolutional network (GCN) or a graph attention network (GAT). Assume the encoder has the following number of layers: ,pass The inter-layer message passing and aggregation mechanism captures the neighborhood association information between task requests and computing resources, and finally outputs encoded features. ,in This is the dimension for encoding features.

[0035] Feature normalization and perceptual semantic extraction: In order to fully utilize the rich perceptual information contained in the encoded features to guide resource state calibration, this embodiment is based on encoded features. Construct a perceptual clustering target. Before inputting the online resource status self-clustering module, first... L2 normalization is performed to project the feature vectors onto a unit hypersphere, ensuring that subsequent dot product-based similarity calculations are not affected by the bias of node feature magnitudes. The specific calculation formula is as follows: in, Representation of the characteristic matrix The first in Node vectors Representing vectors Norm, This is a minimal constant to ensure numerical stability. This operation ensures that subsequent dot product-based similarity calculations are not affected by the bias of node feature magnitudes.

[0036] Sinkhorn Online Resource Status Self-Clustering: The Sinkhorn-Knopp algorithm is used to cluster the normalized resource status. Perform online resource state self-clustering to generate a double random matrix as the semantic-aware matching matrix. The specific steps are as follows: Calculate the normalized features. With a set of learnable resource prototype vectors The similarity between them, and the introduction of a temperature coefficient. Obtain the initial perception assignment matrix The specific implementation process includes: a. Perceptual matching similarity calculation: Define a set of learnable resource prototype vectors ,in The number of potential resource state clusters (e.g., representing implicit resource state prototypes such as "computation-intensive" and "low-latency edge"). Computational normalization features. With resource prototype vector The dot product similarity between them, divided by the temperature coefficient. Scaling by a factor of 0.1 yields the perceptual matching log probability matrix. .

[0037] b. Double random normalization: normalization of the matrix Exponentialization is performed to obtain Then, the Sinkhorn algorithm is used to alternately normalize the rows and columns of the matrix to satisfy the double random constraint (i.e., the sum of each row is 1, and the sum of each column is 1). The iteration rule is as follows: After a few iterations, the matrix Convergence represents the probability distribution of task requests or computing resources belonging to various potential resource state clusters.

[0038] c. Generate perceptual matching targets: Finally, utilize the converged assignment matrix. Calculate the potential association probability between nodes to obtain a value of The semantic-aware matching matrix: The matrix elements in for Continuous values ​​between these ranges represent task requests and computing resource nodes. and The probability of belonging to the same resource state cluster. Compared to only containing The original physical perception record, It can smooth out erroneous mapping associations caused by load fluctuations and complete potential adaptation connections, providing more robust monitoring signals for online branches.

[0039] 2) Calculate the cross-view resource matching prediction matrix using online branches (corresponding to incomplete views). The process includes: using the computing power topology encoder of the online branch to extract features from the incomplete view to obtain online branch encoding features; using the architecture-shared resource regulation decoder to map the target branch encoding features to obtain target branch demand adaptation features; using the architecture-shared resource regulation decoder to map the online branch encoding features to obtain online branch demand adaptation features; performing L2 normalization on the target branch demand adaptation features and the online branch demand adaptation features respectively; calculating the transpose dot product of the normalized online branch demand adaptation features and the normalized target branch demand adaptation features; and performing a dimension softmax operation to obtain the resource matching prediction matrix.

[0040] As a specific implementation of this embodiment, the online branch acts as a "student" in this invention, responsible for reconstructing the potential matching semantics when an incomplete view is input.

[0041] The incomplete view, after being processed with strong perturbation, is input into the computing power topology encoder of the online branch to obtain the encoded features. To reduce the impact of perception measurement noise and achieve cross-view structural alignment, this invention constructs a resource-controlled decoder with shared architecture in both the target branch and the online branch, mapping features to a unified low-dimensional common subspace. By constructing an information bottleneck in this space, the model is guided to automatically filter noise interference during projection. The encoded features of the target branch... Coding features of online branches Each input is fed into its corresponding decoder, and mapped to lower-dimensional demand-adaptive features. and and to and L2 normalization was performed on each part. Then, the normalized online branch requirement adaptation features were calculated. Features that adapt to the requirements of the target branch The transpose of the dot product generates a cross-view resource matching prediction matrix. The specific process is as follows: Network-based Perceptual Association Decoding and Dimensionality Reduction: To reduce the impact of perception measurement noise and achieve cross-view structural alignment, this embodiment maps features to a unified low-dimensional common subspace. By constructing an information bottleneck in this space, the model is guided to automatically filter noise interference. This embodiment constructs a resource-controlled decoder (composed of a multilayer perceptron, or MLP) with shared architecture in both the target branch and the online branch. The encoded features of the target branch are then... Coding features of online branches Each input is fed into its corresponding decoder and mapped to a lower-dimensional feature. and and features and L2 normalization is performed separately. Key point: Feature dimension here. Designed to be smaller than the dimension of the encoded features (For example This "gradual dimensionality reduction" strategy guides the model to discard interference noise during the mapping process, retaining only the most core task requests and computing resource adaptation features.

[0042] Cross-view resource matching prediction matrix calculation: leveraging features of online branches Features of the target branch Interact with the database to generate a cross-view resource matching prediction matrix. The specific calculation formula is as follows: in, The operation is performed along the row dimension of the matrix. This prediction matrix... The goal is to fit the previously generated semantic-aware matching matrix to the greatest extent possible. In other words, the online branch attempts to reconstruct the association mapping between task requests and computing resources perceived by the target branch in a low-dimensional prediction space.

[0043] Step 3: Constructing the resource matching and reconstruction loss function and performing asymmetric parameter updates: Construct a reconstruction loss function based on the semantic-aware matching matrix and the resource matching prediction matrix, update the online branch parameters based on the reconstruction loss function, and apply the converged online branch to the real-time collected computing power network perception data to output the matching relationship between task requests and computing power resources.

[0044] 1) Construction of the loss function for resource matching and reconstruction To suppress random measurement noise in the computing network topology, such as spurious latency caused by network jitter and instantaneous high load fluctuations, this embodiment does not directly use the original physical-aware association records (i.e., the adjacency matrix containing noise) as the supervision signal. Instead, it proposes a semantically-aware guided reconstruction task, utilizing the cross-entropy function to construct the resource matching reconstruction loss. The cross-entropy loss function is used to measure the resource matching prediction matrix. Matching matrix with semantic awareness Differences between them: In the formula, To reconstruct the loss value, This is the resource matching prediction matrix, where N is the number of nodes. This is a semantic-aware matching matrix. The first The and the first The index of each node, where nodes include task request nodes and computing resource nodes.

[0045] By minimizing This training mechanism guides the online branch to predict the associated state perceived by the target branch even when the input is a severely damaged and incomplete view. This allows the model to move beyond blindly fitting incorrect original edges and instead learn to reason about the intrinsic relationship between task requests and computing resources, thereby achieving state calibration and denoising reconstruction for resource regulation.

[0046] 2) Asymmetric parameter update strategy To improve the smoothness and stability during training, this embodiment employs an asymmetric parameter update strategy: Online branch parameter update: Define the network parameters of the online branch as follows (Including computing power topology encoder parameters, resource regulation decoder parameters, and learnable resource prototype vectors) During the backpropagation phase, the loss is calculated. Regarding online branch parameters The gradient is calculated and updated using an optimizer (such as Adam or SGD): in, This is the learning rate.

[0047] Target branch parameter update: Define the network parameters of the target branch as follows The target branch does not accept gradient backpropagation; instead, it updates parameters by following the online branch parameters using an exponential moving average strategy. in, This is the momentum coefficient, and its value range is typically [value range missing]. ,For example This momentum update strategy ensures the stability of the perceptual distribution targets generated by the target branch during training, providing stable perceptual guidance signals for online branches, thereby improving resource regulation.

[0048] Based on the above implementation methods, this invention improves the noise resistance and scheduling accuracy of computing resource regulation. This invention abandons the traditional training method of directly fitting noisy raw sensor records (discrete 0 and 1 hard connections) to the scheduling model, and innovatively utilizes a semantic-aware matching matrix as the reconstruction target. This matrix can filter out interference signals generated by load fluctuations or measurement errors, thereby preventing overfitting of the scheduling model and significantly improving the accuracy of computing resource matching.

[0049] This invention addresses the resource bottleneck in large-scale computing network topologies. To tackle the memory overflow problem in full-graph computation, this invention introduces a topology-preserving subgraph sampling strategy, reducing the dimensionality of complex Sinkhorn computations to the subgraph level. This enables the model to efficiently adapt to computing networks containing a large number of edge nodes, supporting training and inference on large-scale data.

[0050] This invention enables autonomous calibration of sensing data to extract robust features. By projecting coded features into a lower-dimensional latent space using a resource regulation decoder, this invention constructs an effective information bottleneck, guiding the model to filter out interference signals caused by load fluctuations or measurement errors during the regulation process, retaining only robust request task and computing resource adaptation features.

[0051] This invention reduces training overhead and optimizes real-time response efficiency. Unlike contrastive learning methods that rely on negative sampling, this invention maintains training stability through asymmetric momentum updates and cross-entropy reconstruction loss. This design avoids the high computational cost of retrieving negative samples from a massive resource pool, and also avoids mismatches caused by "false negatives" (i.e., misclassifying potential optimal paths as negative samples). While ensuring real-time computing power scheduling, it also improves the robustness of the control scheme.

[0052] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A resource regulation method for semantic awareness guidance and structural denoising in a computing power network, characterized in that, Includes the following steps: Acquire the perception data of task requests and computing resources, and perform topology-preserving subgraph sampling to construct a perception-related subgraph; An asymmetric twin network containing target branches and online branches is constructed based on the perceptual association subgraph; A weak perturbation strategy is applied to the perceptual association subgraph to construct a target view, and a strong perturbation strategy is applied to construct a fragmented view; The target view is input into the target branch to generate a semantic-aware matching matrix. The incomplete view is input into the online branch to extract encoded features. The cross-view resource matching prediction matrix is ​​calculated by obtaining the demand adaptation features of the online branch and the target branch. A reconstruction loss function is constructed based on the semantic perception matching matrix and the resource matching prediction matrix. The online branch parameters are updated based on the reconstruction loss function, and the converged online branch is applied to the real-time collected computing power network perception data to output the control and matching relationship between task requests and computing power resources.

2. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 1, characterized in that, The process of constructing a perceptual association subgraph includes: Based on the acquired task request and computing resource perception data, an original perception association graph is constructed, and an enhanced adjacency matrix is ​​obtained by adding self-loops to the adjacency matrix of the original perception association graph. After feature normalization of the enhanced adjacency matrix, a graph partitioning algorithm is used to partition the entire set of nodes of the original perceptual association graph with the feature normalization to obtain several clusters. The set of nodes in the subgraph is determined based on a subset of clusters randomly selected in each training iteration; A perceptual association subgraph is obtained based on the set of subgraph nodes.

3. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 1, characterized in that, The weak perturbation strategy includes setting a first feature mask rate and a first association mapping edge deletion rate; The strong perturbation strategy includes setting a second feature mask rate and a second association mapping edge deletion rate; The second feature mask rate is greater than the first feature mask rate, and the second association mapping edge deletion rate is greater than the first association mapping edge deletion rate.

4. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 1, characterized in that, The process of generating a semantically aware matching matrix by inputting the target view into the target branch includes: The target branch encoding features are obtained by using a computing power topology encoder for the target branch to extract features from the target view. The target branch coding features are subjected to L2 normalization to obtain normalized coding features; The Sinkhorn-Knopp algorithm is used to perform online resource state self-clustering on the normalized encoded features and the learnable resource prototype vector to obtain the probability distribution matrix. The semantic-aware matching matrix is ​​obtained by calculating the potential association probability between nodes based on the probability distribution matrix.

5. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 4, characterized in that, The process of using the Sinkhorn-Knopp algorithm to perform online resource state self-clustering on the normalized encoded features and learnable resource prototype vectors to obtain the probability distribution matrix includes: The perceptual matching log probability matrix is ​​obtained by scaling the dot product similarity between the normalized encoded features and the learnable resource prototype vector with a temperature coefficient. The initial perception assignment matrix is ​​obtained by exponentializing the perception matching log probability matrix. The initial sensing allocation matrix is ​​subjected to alternating row normalization and column normalization operations until convergence, resulting in a probability distribution matrix.

6. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 4, characterized in that, The process of inputting the incomplete view into the online branch to extract encoded features, and calculating the cross-view resource matching prediction matrix by obtaining the demand adaptation features of the online branch and the target branch, includes: An online branching computational topology encoder is used to extract features from the incomplete view to obtain online branching coding features; The target branch requirement adaptation features are obtained by mapping the target branch encoding features using a shared architecture resource regulation decoder. A resource regulation decoder with a shared architecture is used to map the online branch encoding features to obtain online branch demand adaptation features; The target branch demand adaptation features and online branch demand adaptation features are both subjected to L2 normalization. The transpose dot product of the normalized online branch demand adaptation features and the normalized target branch demand adaptation features is calculated and then subjected to a dimension softmax operation to obtain the resource matching prediction matrix.

7. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 6, characterized in that, The resource regulation decoder is composed of a multi-layer sensing mechanism.

8. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 1, characterized in that, The reconstruction loss function is: ； In the formula, To reconstruct the loss value, This is the resource matching prediction matrix, where N is the number of nodes. This is a semantic-aware matching matrix. The first The and the first Index of each node.

9. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 8, characterized in that, The process of updating the online branch parameters based on the reconstructed loss function includes: The gradient with respect to the online branch parameters is calculated based on the reconstruction loss function; The optimized online branch parameters are updated by gradient updating the online branch parameters.

10. The resource regulation method for semantic awareness guidance and structural denoising in a computing network according to claim 9, characterized in that, An exponential moving average strategy is used to process the updated online branch parameters to obtain the updated target branch parameters. The update expression for the target branch parameters is as follows: ； In the formula, The momentum coefficient, For online branch parameters, For the updated target branch parameters, The parameters for the target branch before the update.

Images