Search

CN-121981109-A - Personalized graph federal learning method and system based on attention and structural perception

CN121981109ACN 121981109 ACN121981109 ACN 121981109ACN-121981109-A

Abstract

The invention provides a personalized graph federal learning method and system based on attention and structural perception, and relates to the technical field of federal learning and graph data processing. Aiming at core pain points with deficiency of edges, data heterogeneity and insufficient global semantic capture of a cross-client side in federal sub-graph scenes such as cross-institution scientific research collaboration, e-commerce platform collaboration and biological network joint analysis, the invention adopts a four-step architecture of three-view attention semantic extraction, structure perception supplementation, clustering prototype comparison and personalized aggregation, namely, three-view attention is guided through a neighbor center, global sparsity and a prototype, local association, global long-distance dependence and category semantic anchor points are respectively captured, topological structure features are supplemented through a graph-convolution network, data heterogeneity is relieved based on clustering-prototype federal comparison learning, and personalized parameter aggregation is realized through prototype similarity. The method is suitable for actual tasks such as document classification, commodity recommendation, protein function prediction and the like.

Inventors

  • HUANG WEI
  • ZHU CHAO

Assignees

  • 福州大学

Dates

Publication Date
20260505
Application Date
20260131

Claims (10)

  1. 1. The personalized graph federation learning method based on the attention and the structure perception is characterized by realizing node characterization learning through cooperation of triple-view attention perception, structure supplementation and personalized aggregation, wherein the personalized graph federation learning method based on the attention and the structure perception further comprises the following contents: step S1, global and local collaborative initialization, which comprises the steps that a server acquires global configuration information of a cross-mechanism federation scene, a client loads local sub-image data and initializes model parameters, and a secure communication link between the client and the server is established; step S2, triple-view attention semantic extraction, which comprises a triple-view attention module formed by a client-side through neighbor center attention, global sparse attention and prototype guiding attention, and respectively capturing local neighbor consistency, global long-distance dependence and category semantic anchor point information to generate multi-dimensional semantic representation; s3, supplementing structural perception features, wherein the structural perception features comprise that a client extracts local sub-graph topological structure information through a graph convolution network, and the local sub-graph topological structure information is fused with multi-dimensional semantic representation to obtain node comprehensive representation with both semantics and structure; step S4, cluster-prototype federation contrast optimization, which comprises the steps that a client calculates a local category prototype and uploads the local category prototype to a server, the server generates a cluster prototype through K-means clustering and issues the cluster prototype, and the client calculates a mixed loss update local model by combining the cluster prototype; Step S5, prototype similarity personalized aggregation comprises the steps that the server calculates similarity weights based on prototypes of all clients, personalized model parameters are generated through weighted aggregation and sent down, and the clients update models to complete one round of training.
  2. 2. The personalized graph federal learning method based on attention and structural awareness according to claim 1, wherein step S1 comprises the following: In step S1, global configuration information of a cross-organization federal scene comprises global configuration information of scientific research institutions, electronic commerce platforms and biomedical treatment, wherein the global configuration information comprises the number of clients and the total number of data categories, and the client loading local sub-image data comprises a scientific research literature network, a commodity co-purchase network and a protein interaction network.
  3. 3. The personalized graph federal learning method based on attention and structural awareness according to claim 1, wherein step S2 comprises the following: step S21, neighbor center attention extraction, namely sampling 1-order or 2-order neighbors of a target reading node according to actual tasks such as scientific research literature classification, E-commerce commodity clustering, protein function prediction and the like, and constructing a sequence Wherein As an original characteristic of the node, A neighbor sampling function; Step S22, global sparse attention extraction, namely aiming at the problem of global information fracture caused by cross-mechanism sub-graph edge deletion, uniformly sampling partial nodes from a local sub-graph, and constructing a global sequence: ; Wherein the method comprises the steps of (. Cndot.) is a uniform sampling function that complements global long-range semantic dependencies.
  4. 4. A personalized graph federal learning method based on attention and structural awareness according to claim 3, wherein step S2 further comprises: Step S23, semantic characterization fusion, namely combining the two sequences to obtain the sequence: ; outputting semantic characterization through transducer multi-head attention calculation The calculation process is as follows: ; ; ; wherein Q, K, V are respectively composed of input sequences By a learning projection matrix The linear transformation is used for subsequent attention calculation; Then, in order to avoid unstable gradient caused by overlarge dot product value, the attention score is divided by a scaling factor d before Softmax, wherein d is the hidden layer dimension; Subsequently, each attention head outputs Splicing and linearly transforming again to obtain final characterization fusing global and local semantics ; Step S24, prototype guided attention extraction, for cross-client data heterogeneity problem, based on Computing local category prototypes Wherein Representing a subset of nodes in a client sub-graph marked as class m, wherein As the mass center of class M in the common feature space, all M classes of prototypes are arranged in sequence and then are represented by the current reading node Splicing to obtain a prototype sequence: ; wherein the prototype sequence is then fed into a multi-headed attention module Enabling nodes to explicitly interact with each prototype type and output Through this process, node features are injected into category semantic anchors across clients, alleviating semantic drift due to topology breaks.
  5. 5. The personalized graph federal learning method based on attention and structural awareness according to claim 1, wherein step S3 comprises the following: The step S31 is that the content supplemented by the structural perception feature comprises the following steps: extracting local sub-graph topology information by adopting graph rolling network, and generating structural representation through message passing and node updating mechanism The calculation formula is: ; Wherein the method comprises the steps of Representing a first-order set of neighbors of the readout node r, Is responsible for passing neighbor information along the edges and aggregating, Then fusing the aggregation result with the characteristics of the node itself, and finally outputting Namely carrying structural perception representation of local topology induction bias; the node comprehensive characterization is obtained by fusing semantic and structural information according to the following formula: , Wherein the method comprises the steps of And classifying the prediction result for the node.
  6. 6. The personalized graph federal learning method based on attention and structural awareness according to claim 1, wherein step S4 comprises the following: Step 41, uploading local prototypes, namely uploading each prototype to a server by a client, avoiding privacy disclosure caused by original data transmission, and meeting privacy regulation requirements of cross-institution data sharing; Wherein the prototype expression is: ; wherein M is the number of all classes; step S42, mixed loss calculation, including the calculation client side calculating cross entropy loss and prototype comparison loss, Wherein the client calculates the cross entropy loss as: ; Wife prototype contrast loss of ; Wherein the method comprises the steps of A feature vector representing a class m of the fused node q, For the j-th clustering prototype with m category issued by the server, tau >0 is the temperature coefficient, controlling the sharpness of the contrast distribution, N is the total number of nodes, and finally, by the balance coefficient Fusion to obtain total loss And (5) reinforcing class boundary distinction.
  7. 7. The personalized graph federal learning method based on attention and structural awareness according to claim 1, wherein step S5 comprises the following: Step S51, prototype similarity calculation, which comprises the step that the server calculates the sum of all types of prototype similarity of the client k and the client t according to the types: Step S52, personalized parameter generation, including the server calculating weight matrix according to prototype similarity Wherein Representing normalized aggregate weights between client k and client t, indicating that the larger the aggregate weights indicate the closer the two client data distributions are, and re-using the weights to model parameters Weighted summation is carried out, and the server calculates aggregation parameters of the client k ; Wherein the aggregation parameters are used to avoid local task performance loss caused by unified aggregation.
  8. 8. The personalized graph federal learning method based on attention and structural awareness of claim 7, wherein step S5 further comprises: And step 53, repeating the steps S2-S5 to a preset training round, and outputting a personalized federation diagram converter model adapting to the federation sub-graph scene for predicting and other tasks.
  9. 9. A personalized graph federal learning system based on attention and structural awareness, comprising an electronic device, wherein the electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a personalized graph federal learning method based on attention and structural awareness as claimed in any one of claims 1 to 8 when executing the computer program.
  10. 10. A personalized graph federation learning system based on attention and structural awareness, comprising a computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements a personalized graph federation learning method based on attention and structural awareness as claimed in any one of claims 1 to 8.

Description

Personalized graph federal learning method and system based on attention and structural perception Technical Field The invention provides a personalized graph federal learning method and system based on attention and structural perception, and relates to the technical fields of federal learning, graph transformation and node classification. Background Federal subgraph learning is a key technology for solving the problem of data privacy protection and collaborative modeling of cross-organization graphs, and is widely applied to the fields of scientific research literature analysis, electronic commerce recommendation systems, biological information mining and the like. Under the scene, the global graph is divided into a plurality of sub-graphs which are scattered at different clients (such as different scientific research institutions, different electronic commerce platforms and different hospitals), and two core pain points exist, namely, the sub-graphs are broken due to the fact that edges of the clients are deleted, for example, documents of an A institution have no introduction association of an B institution, commodities of an A platform have no co-purchase data of a B platform, global semantic information is difficult to capture, data distribution heterogeneity among the clients is obvious, for example, document theme preference difference of different scientific research institutions and commodity class distribution difference of different electronic commerce platforms exist, and node overlapping (such as document sharing and user sharing across institutions) exists, and parameter conflict and performance reduction are easily caused by a traditional federal aggregation strategy. The existing federal graph learning method has obvious defects that the graph convolutional network-based method depends on complete edge connection, information transmission is interrupted when edges are missing, an overcorrection problem is easily generated, a cross-mechanism cooperation scene is difficult to adapt, the graph transform-based method depends on neighbor sampling, local information can only be obtained, global semantic guidance is lacking, information breakage caused by the cross-client edge missing cannot be solved, a single global prototype is mostly adopted in the prototype learning method, client distribution difference is ignored, scenes with strong heterogeneity such as a power supplier and scientific research are difficult to adapt, an individualized aggregation strategy lacks an effective semantic similarity measurement basis, and global alignment and local task adaptation (such as document classification preference of a scientific research institution and commodity recommendation trend of an electric supplier platform) are difficult to balance. Therefore, on the premise of protecting the inter-mechanism data privacy, the global semantic and local structure information is completely captured, the heterogeneity of the client is relieved, and the method becomes a key challenge in the federal sub-graph learning field. ‌ A Disclosure of Invention In view of the above, in order to make up for the blank and the deficiency of the prior art, the invention provides a personalized graph federal learning method and system based on attention and structure perception, which are used for solving the problems of incomplete semantic capture, poor client heterogeneity adaptation, weak model generalization capability and the like in the prior art. The invention provides a personalized graph federal learning method and a personalized graph federal learning system based on attention and structural perception, which comprise the following contents: the invention provides a personalized graph federation learning method based on attention and structure perception, which is characterized in that node characterization learning is realized through cooperation of triple-view attention perception, structure supplementation and personalized aggregation, wherein the personalized graph federation learning method based on the attention and structure perception further comprises the following contents: step S1, global and local collaborative initialization, which comprises the steps that a server acquires global configuration information of a cross-mechanism federation scene, a client loads local sub-image data and initializes model parameters, and a secure communication link between the client and the server is established; step S2, triple-view attention semantic extraction, which comprises a triple-view attention module formed by a client-side through neighbor center attention, global sparse attention and prototype guiding attention, and respectively capturing local neighbor consistency, global long-distance dependence and category semantic anchor point information to generate multi-dimensional semantic representation; s3, supplementing structural perception features, wherein the structural perception features comprise that a client e