CN-122027553-A - Analysis scheduling method and system based on multi-label routing

CN122027553ACN 122027553 ACN122027553 ACN 122027553ACN-122027553-A

Abstract

The invention discloses an analysis scheduling method and system based on multi-label routing, the method comprises the steps of receiving multi-label hypergraph data, performing Take decomposition on hypergraph adjacent tensors, applying sparse constraint, extracting multi-label dependent structures of cores, generating a low-rank factor matrix, extracting a hyperedge embedded vector according to the low-rank factor matrix, compressing the hyperedge embedded vector into binary hash codes through a local sensitive hash module, splicing the binary hash codes with residual floating point features to form a mixed hyperedge embedded vector, calculating attention weight through a multi-label compatibility function for nodes in a graph neural network based on the mixed embedded vector, performing neighborhood aggregation, updating node states through a gating circulation unit to obtain node embedded representation, generating the node embedded representation through a routing scheme through a path searching algorithm, evaluating different schemes through a path scoring function, and outputting an optimal scheduling route according to a scoring result. The invention improves the calculation and message transmission efficiency of the large-scale multi-label graph.

Inventors

LU YONGLING
HU YANJIE
HU CHENGBO
LIU JIANJUN
MIAO YANG
ZHANG CAN
XIN SHUAIKUI
LIU ZHENGYU
XUE HAI
LIU ZIQUAN

Assignees

国网江苏省电力有限公司电力科学研究院
国网江苏省电力有限公司
江苏省电力试验研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (8)

1. An analytical scheduling method based on multi-label routing is characterized by comprising the following steps: s10, receiving multi-label hypergraph data, performing Take decomposition on hypergraph adjacent tensors, applying sparse constraint, extracting a multi-label dependent structure of a core, and generating a low-rank factor matrix; S20, extracting an over-edge embedded vector according to the low-rank factor matrix, compressing the over-edge embedded vector into a binary hash code through a learnable local sensitive hash module, and splicing the binary hash code with residual floating point characteristics to form a mixed over-edge embedded vector; s30, for a target node in the graph neural network, calculating attention weight through a multi-label compatibility function based on the mixed embedding vector, performing efficient neighborhood aggregation, and updating the node state through a gating circulation unit to obtain node embedding representation containing neighborhood multi-label information; S40, generating a routing scheme by embedding the node representation through a path search algorithm, evaluating service quality indexes of different routing schemes through a path scoring function, and outputting the optimal data flow routing and scheduling scheme according to the scoring result.
2. The analytical scheduling method based on multi-label routing according to claim 1, wherein the steps of receiving multi-label hypergraph data, performing tak decomposition on the hypergraph adjacent tensor and applying sparse constraint, extracting multi-label dependency structure of a core, and generating a low-rank factor matrix comprise the following steps: receiving multi-label hypergraph data, and decomposing the hypergraph adjacent tensor into a core tensor, nodes, superedges and label three factor matrixes; Applying a structured sparsity constraint to the hypergraph adjacency tensor to preserve core connections associated with the multi-label route; And solving an optimization problem with sparse constraint by using an iterative algorithm, and outputting a final sparse factor matrix.
3. The analytical scheduling method based on multi-label routing according to claim 2, wherein the optimization problem with sparse constraint is solved by using an iterative algorithm, and a final sparse factor matrix is output, and the method specifically comprises the following sub-steps: In each iteration, other variables are fixed first, and an original factor matrix and a core tensor are updated by solving a least square problem; Independently updating auxiliary variables by using a soft threshold operator to realize sparsification; And updating the dual variables according to the current constraint residual until the original residual and the dual residual are converged to a preset threshold value, and finally obtaining sparse node, superb and label factor matrixes.
4. The analytical scheduling method based on multi-label routing according to claim 1, wherein the method comprises the steps of compressing the superside embedded vector into binary hash codes by a learnable locality sensitive hash module, and splicing the binary hash codes with residual floating point characteristics to form a mixed superside embedded vector, and the method comprises the following steps: Inputting the superside embedded vector into a local sensitive hash module, mapping the superside embedded vector into a real space by the module, and binarizing by applying a symbol function to generate a compact binary hash code; The over-edge embedded vector generates a residual floating point feature vector for compensating information loss in the hash process through a residual projection matrix, and the dimension of the residual feature can be distributed in a self-adaptive manner according to the importance level of the over-edge; And splicing the generated binary hash code with the residual floating point characteristic to obtain the mixed superside embedded vector.
5. The analytical scheduling method based on multi-label routing according to claim 4, wherein the importance ranking of the superside is specifically divided into the following sub-steps: determining and collecting superside related indexes and specific data about importance grade division, and setting scoring standards for the indexes; And calculating comprehensive scores and dividing the supersides into different grades according to the comprehensive score results.
6. The analytical scheduling method based on multi-label routing according to claim 1, wherein for a target node in a graph neural network, attention weights are calculated through a multi-label compatibility function based on mixed embedding vectors, efficient neighborhood aggregation is performed, node states are updated through a gating circulation unit, and node embedding representations containing neighborhood multi-label information are obtained, and the method specifically comprises the following sub-steps: performing high-efficiency attention weight calculation through a multi-label compatibility function according to the mixed embedded vector, the node self state and the label vector; generating an actual transfer message by using residual floating point characteristics in the mixed embedded vector, and carrying out node message aggregation according to the actual transfer message and the attention weight; and updating the information of each node by using the gating circulating unit as an updating function to obtain the node embedded representation containing the neighborhood multi-label information.
7. The analytical scheduling method based on multi-label routing according to claim 1, wherein the node embedding representation is generated by a routing scheme through a path search algorithm, the quality of service indexes of different routing schemes are evaluated through a path scoring function, and the optimal data flow routing and scheduling scheme is output according to the scoring result, and the method specifically comprises the following sub-steps: Dividing a scheduling period into a plurality of time slots, acquiring node embedded representations for each time slot, decomposing the node embedded representations, calculating edge cost according to a decomposition vector, and constructing a time expansion graph; sorting the edge cost as a cost set, searching a plurality of potential feasible paths from a source node to a target node through a path searching algorithm based on the time expansion graph and the cost set, and obtaining a scheduling scheme sequence; And evaluating the service quality index of the scheme in the scheduling scheme sequence according to a preset evaluation standard by a path scoring function, and outputting the optimal data flow route and scheduling scheme according to the scoring result.
8. An analytical dispatch system based on multi-label routing, comprising: The hypergraph input module is used for receiving the multi-label hypergraph data, performing Take decomposition on the hypergraph adjacent tensor, applying sparse constraint, extracting a multi-label dependent structure of a core, and generating a low-rank factor matrix; the extraction and splicing module extracts the superside embedded vector according to the low-rank factor matrix, compresses the superside embedded vector into a binary hash code through the learnable local sensitive hash module, and splices the binary hash code with the residual floating point characteristic to form a mixed superside embedded vector; The aggregation updating module is used for calculating attention weights of target nodes in the graph neural network through a multi-label compatibility function based on the mixed embedding vectors, carrying out efficient neighborhood aggregation, and updating node states through a gating circulating unit to obtain node embedding representations containing neighborhood multi-label information; and the scheme generating and evaluating module is used for generating the routing scheme by embedding the node representation through a path searching algorithm, evaluating the service quality indexes of different routing schemes through a path scoring function and outputting the optimal data flow routing and scheduling scheme according to the scoring result.

Description

Analysis scheduling method and system based on multi-label routing Technical Field The invention relates to the field of Internet, in particular to an analysis scheduling method and system based on multi-label routing. Background With the wide application of multi-label graph data in the fields of communication networks, bioinformatics and the like, efficient modeling of complex multi-label dependency relationships in hypergraphs becomes a key challenge. The traditional graph neural network captures a topological structure through message transmission among nodes, but the design of the traditional graph neural network is based on binary edge connection of a common graph, and high-order correlation of superedges is difficult to directly process. In recent years, the hypergraph neural network partially solves the problem by expanding a message passing mechanism to the hyperedges, but the computational complexity of the hypergraph neural network increases exponentially with the number of the hyperedges, particularly when processing large-scale multi-label routing tasks, the hypergraph neural network faces remarkable memory and computational bottlenecks, and the existing hypergraph message passing layer generally statically aggregates neighborhood information and lacks a response mechanism for dynamic change of routing states. Based on the analysis scheduling method, the invention provides an analysis scheduling method based on multi-label routing. Disclosure of Invention The invention provides an analysis scheduling method based on multi-label routing, which is characterized by comprising the following steps: s10, receiving multi-label hypergraph data, performing Take decomposition on hypergraph adjacent tensors, applying sparse constraint, extracting a multi-label dependent structure of a core, and generating a low-rank factor matrix; S20, extracting an over-edge embedded vector according to the low-rank factor matrix, compressing the over-edge embedded vector into a binary hash code through a learnable local sensitive hash module, and splicing the binary hash code with residual floating point characteristics to form a mixed over-edge embedded vector; s30, for a target node in the graph neural network, calculating attention weight through a multi-label compatibility function based on the mixed embedding vector, performing efficient neighborhood aggregation, and updating the node state through a gating circulation unit to obtain node embedding representation containing neighborhood multi-label information; S40, generating a routing scheme by embedding the node representation through a path search algorithm, evaluating service quality indexes of different routing schemes through a path scoring function, and outputting the optimal data flow routing and scheduling scheme according to the scoring result. The analysis scheduling method based on the multi-label routing is characterized by receiving multi-label hypergraph data, performing Take decomposition on the hypergraph adjacent tensor, applying sparse constraint, extracting a multi-label dependent structure of a core, and generating a low-rank factor matrix. The method comprises the following steps: receiving multi-label hypergraph data, and decomposing the hypergraph adjacent tensor into a core tensor, nodes, superedges and label three factor matrixes; Applying a structured sparsity constraint to the hypergraph adjacency tensor to preserve core connections associated with the multi-label route; And solving an optimization problem with sparse constraint by using an iterative algorithm, and outputting a final sparse factor matrix. According to the analysis scheduling method based on the multi-label routing, an iterative algorithm is used for solving the optimization problem with the sparse constraint, and the final sparse factor matrix is output. The method comprises the following steps: In each iteration, other variables are fixed first, and an original factor matrix and a core tensor are updated by solving a least square problem; Independently updating auxiliary variables by using a soft threshold operator to realize sparsification; And updating the dual variables according to the current constraint residual until the original residual and the dual residual are converged to a preset threshold value, and finally obtaining sparse node, superb and label factor matrixes. According to the analysis scheduling method based on the multi-label routing, the super-edge embedded vector is compressed into the binary hash code through the learnable local sensitive hash module, and the binary hash code is spliced with the residual floating point characteristic to form the mixed super-edge embedded vector. The method comprises the following steps: Inputting the superside embedded vector into a local sensitive hash module, mapping the superside embedded vector into a real space by the module, and binarizing by applying a symbol function to generate a compact binary hash code;