Heterogeneous task-oriented big language model enhanced federal map learning method and system

Abstract

The invention discloses a heterogeneous task-oriented big language model enhanced federal map learning method and system, and the method comprises the steps: constructing a product manifold containing hyperbolic, spherical and Euclidean space, and guiding the dynamic adjustment of an edge weight through Ollivier-Ricci curvature, so that a big language model is subjected to the dynamic adjustment of an edge weight under the condition that the annotation data is rare; and complex hierarchical, cyclic and flat geometric structures in the graph data can still be fully captured, so that the deep understanding and semantic characterization capability of the model on the graph topological information can be obviously enhanced. In order to solve the cross-client task heterogeneous problem, a common feature subspace with geometric consistency is extracted from heterogeneous parameter updating through an innovative shared subspace aggregation mechanism by utilizing curvature weighted whitening and two-stage singular value decomposition technologies, so that semantic interference peculiar to tasks can be effectively filtered out, and the task heterogeneous performance is improved. The efficient and robust aggregation of model parameters under different learning targets is realized, and finally the generalization performance of a global model in a heterogeneous task environment is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of heterogeneous task processing combined with big data, and particularly relates to a large language model enhanced federated graph learning method and system for heterogeneous tasks. BACKGROUND

[0002] Graph structure data widely exists in various application scenarios in the real world, such as user-goods interaction networks in e-commerce, social relationship networks in social media, etc. In order to fully utilize the data scattered in different organizations or devices for model training under the premise of meeting the data privacy requirements of privacy protection regulations, federated graph learning emerges as a new paradigm. Federated graph learning allows participants to collaboratively train a global model by exchanging model parameters or gradients without sharing the original graph data, thereby breaking the data silos.

[0003] In recent years, large language models have begun to be introduced into the field of graph learning due to their excellent generalization ability and reasoning ability in natural language processing tasks. Existing combination methods mainly fall into two categories: one is to use large language models as enhancers to enrich node features with semantic knowledge; the other is to directly use large language models as predictors by converting graph structure and node information into Token sequences to input into large models for prediction. This combination provides new opportunities for processing complex graph data, especially in utilizing the powerful pre-training knowledge of large models to improve the generalization of graph models.

[0004] However, existing federated graph learning methods still face severe challenges in practical applications. First, the task heterogeneity problem. Existing federated graph learning methods usually assume that all clients are performing the same graph learning task. But in real-world scenarios, the business goals of different clients may be completely different, for example, some clients focus on node classification such as user portrait inference, while others focus on link prediction such as friend recommendation. This difference in task objectives leads to significant heterogeneity in the semantic space and optimization direction of parameter updates of each client model, and traditional parameter averaging aggregation strategies cannot effectively integrate these heterogeneous updates, resulting in a decline in the performance of the global model.

[0005] Secondly, there is a lack of data and limitations in geometric expression. In the federated learning setting, a single client usually only has a small number of labeled samples. At the same time, existing large model enhanced graph learning methods are mostly limited to using Euclidean geometric space to represent graph data. This approach ignores the complex non-Euclidean geometric characteristics of graph data, making it difficult for the model to capture deep graph structure information when client data is limited, thereby limiting its reasoning ability. Therefore, how to break through the limitations of a single geometric space in a data-scarce and task-heterogeneous federated environment and achieve efficient and consistent aggregation of heterogeneous parameters under different learning goals is a key technical problem that needs to be solved in the current federated graph learning field.

[0006] Current federated graph learning methods have made significant progress in addressing data privacy and distribution heterogeneity, but there are still obvious limitations in dealing with task heterogeneity and data scarcity across clients.

[0007] On the one hand, the solution to heterogeneity mainly focuses on the differences in resource or data distribution. For example, although the patent application with application number CN202210823547.6 proposes "heterogeneous task requirements", its core is to balance training time and model accuracy, addressing resource-level heterogeneity. The patent application with application number CN202210499734.3 alleviates the problem of non-independent and identically distributed data through personalized supplementary models. However, these methods usually presuppose that all participants perform the same graph learning task, such as node classification, and do not consider the functional task heterogeneity that exists in reality, such as some clients performing node classification and others performing link prediction. The fundamental differences in task objectives can lead to conflicts in the optimization direction and semantic space of model parameters, making it difficult for traditional weighted aggregation to extract cross-task common knowledge.

[0008] On the other hand, attempts to align features and domains have introduced manifold or feature space construction techniques. The patent application with application number CN202510947388.4 aligns heterogeneous features by constructing a shared feature space through random walk; the patent application with application number CN202410769913.3 uses manifold learning to solve the heterogeneity of multi-domain image data. However, these methods mainly address feature space or data domain shifts and rely heavily on Euclidean geometric assumptions or complex clustering reconstruction, lacking deep adaptation to complex geometric structures of graph data such as trees and rings.

[0009] In addition, although existing aggregation mechanisms consider similarity weighting, they still operate in the original parameter space or simple feature space, lacking a perfect extraction mechanism, making it difficult to accurately separate geometrically consistent common features and task-specific private features when the dimensions of heterogeneous task parameters are uneven and the variances are different.

[0010] In summary, the prior art faces the core bottlenecks of multi-task collaborative optimization conflict, insufficient understanding of graph structure information under few samples, and limited single geometric space representation when facing real heterogeneous task scenarios. SUMMARY

[0011] In view of the above, the purpose of the present application is to provide a large language model enhanced federated graph learning system for heterogeneous tasks, which enhances the deep understanding of the graph structure of the large language model by constructing a multi-manifold space, and extracts common features across tasks using a shared subspace mechanism, thereby realizing the consistency aggregation of heterogeneous parameters.

[0012] To achieve the above-mentioned purpose of the application, an embodiment provides a large language model enhanced federated graph learning system for heterogeneous tasks, comprising a server and a plurality of clients holding local private graph data; The client is configured to project the local private graph data into a semantic token sequence through a projector, specifically comprising: constructing a product manifold comprising hyperbolic, spherical and Euclidean spaces and calculating the node features of the graph data projected into the product manifold, using Ollivier-Ricci curvature as a geometric prior to dynamically adjust the edge weight, and using the edge weight to perform geometric aggregation on the node features on the product manifold to realize geometric perception propagation and then map back to the Euclidean tangent space to obtain the semantic token sequence; using the local task learning of the large language model based on the semantic token sequence to update the projector parameter and the curvature importance vector, and sending the projector parameter and the calculated curvature importance vector to the server; The server is configured to realize projector parameter aggregation using a shared subspace aggregation mechanism, specifically comprising: constructing a global parameter update matrix based on the projector parameters of all clients, and calculating the parameter update matrix after whitening the curvature importance vector, performing two-stage SVD decomposition on the projector parameters of all clients based on the whitened parameter update matrix to extract a shared geometric subspace, and projecting the projector parameter update amount of each client using the extracted shared geometric subspace, then obtaining a global update amount by weighted aggregation and realizing global parameter update; the globally updated projector parameters are distributed to each client for the next round of federated learning.

[0013] A large language model enhanced federated graph learning method for heterogeneous tasks, the method uses the above-mentioned system, comprising the following steps: The client projects local private graph data into a semantic token sequence through a projector, specifically including: constructing a product manifold containing hyperbolic, spherical and Euclidean spaces and calculating node features of the graph data projected into the product manifold, taking Ollivier-Ricci curvature as a geometric prior to dynamically adjust edge weights, and using the edge weights to perform geometric aggregation on the node features in the product manifold to realize geometric perception propagation and then map back to the Euclidean tangent space to obtain the semantic token sequence; and updating projector parameters and calculating curvature importance vectors based on local task learning of the semantic token sequence by a large language model, and sending the projector parameters and the curvature importance vectors to the server; The server aggregates the projection parameters using a shared subspace aggregation mechanism, specifically including: constructing a global parameter update matrix based on the projector parameters of all clients, and calculating a parameter update matrix after the curvature importance vector is weighted and whitened, performing two-stage SVD decomposition on the whitened parameter update matrix to extract a shared geometric subspace based on the projector parameters of all clients, and projecting the projector parameter update amount of each client using the extracted shared geometric subspace, and then obtaining a global update amount by weighted aggregation and realizing global parameter update; and downloading the globally updated projector parameters to each client for use in the next round of federated learning.

[0014] Compared with the prior art, the present application has at least the following beneficial effects: By constructing a product manifold containing hyperbolic, spherical and Euclidean spaces, and using Ollivier-Ricci curvature to guide the dynamic adjustment of edge weights, the large language model can still fully capture the complex hierarchical, cyclic and flat geometric structures in the graph data even in the case of insufficient client labeled data, thereby significantly enhancing the model's deep understanding and semantic representation ability of the graph topology information.

[0015] In order to solve the task heterogeneity problem across clients, an innovative shared subspace aggregation mechanism is used to extract a geometrically consistent public feature subspace from the heterogeneous parameter updates by using curvature weighted whitening and two-stage singular value decomposition technology. This mechanism can effectively filter out task-specific semantic interference, realize efficient and robust aggregation of model parameters under different learning goals, and ultimately improve the generalization performance of the global model in a heterogeneous task environment. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below of the drawings needed in the embodiments or prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0017] Figure 1 is a structure and flow diagram of a large language model enhanced federated graph learning system for heterogeneous tasks provided by the embodiment; Figure 2 is a flow diagram of a large language model enhanced federated graph learning method for heterogeneous tasks provided by the embodiment. DETAILED DESCRIPTION

[0018] In order to make the object, technical scheme and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the protection scope of the present application.

[0019] The inventive concept of the present application is to solve the technical problems of poor model generalization ability and low collaborative training efficiency of existing federated graph learning methods when facing client task heterogeneity and data scarcity. Existing methods usually preset all clients to perform uniform learning tasks, which is difficult to adapt to the actual scenario where heterogeneous tasks such as node classification and link prediction coexist, resulting in parameter semantic conflict when aggregating global model, and limited generalization. At the same time, in the typical few-shot scenario of federated learning, existing large language model enhancement schemes are mostly limited to single Euclidean space projection, which cannot fully capture the complex internal geometric structure of graph data, resulting in insufficient understanding of the model to the graph topology information.

[0020] Therefore, the present application proposes a large language model enhanced federated graph learning scheme for heterogeneous tasks, the core idea of which is geometric enhancement and subspace alignment. The specific process is as follows: First, on the client side, the present application designs a projector. Instead of being limited to a single Euclidean space, the projector constructs a product manifold containing hyperbolic, spherical and Euclidean spaces, uses Ollivier-Ricci curvature as a geometric prior to dynamically adjust edge weights, and maps complex graph topology structure to semantic Token that can be understood by large language model, thereby enhancing the model's representation ability for graph structure under few-shot conditions.

[0021] Secondly, on the server side, the present application designs a shared subspace aggregation mechanism. For the heterogeneous parameter updates submitted by each client due to different tasks, the server uses the curvature importance vector to whiten the parameters, and extracts the shared geometric subspace across tasks through two-stage singular value decomposition. The aggregation in this subspace can effectively filter the noise interference specific to the task and retain the geometric consistent common features. Finally, the server distributes the aggregated global projector parameters back to the client, iterates cyclically until the model converges, and realizes efficient and robust collaborative training of heterogeneous tasks.

[0022] As Figure 1As shown, the heterogeneous task-oriented large language model enhanced federated graph learning system provided by the embodiment includes a server and multiple clients holding local private graph data. The client projects the local private graph data into a semantic token sequence through a projector, updates the projector parameters and calculates the curvature importance vector based on the local task learning of the semantic token sequence by the large language model, and sends the projector parameters and the curvature importance vector to the server; the server uses a shared subspace aggregation mechanism to realize projector parameter aggregation, and distributes the updated global projector parameters to each client for the next round of federated learning.

[0023] At each client , given the local private graph data , where is a node feature matrix, is an edge set, is a node set. The private graph data includes a user-goods interaction network, where nodes represent users and goods, edges represent purchase relationships between users and goods, association relationships between goods, and friend relationships between users.

[0024] To solve the problem that a single Euclidean space cannot fully capture the complex geometric structure of a graph, the local private graph data is projected into a semantic token sequence using a projector, specifically including: (1) First, construct a product manifold composed of three complementary geometric spaces and calculate the node features of the graph data projected onto the product manifold.

[0025] Define the product manifold as the Cartesian product of hyperbolic space , spherical space , and Euclidean space : wherein represents the hyperbolic curvature, used to capture hierarchical structure; represents the spherical curvature, used to capture cyclic structure; is the intrinsic dimension of each independent geometric space component when constructing the product manifold; map the original node features in the graph data to the tangent space of each manifold through a manifold-specific encoder , and then project them onto the manifold using exponential mapping to obtain the node features projected onto the product manifold, expressed by the formula as: wherein represents the exponential mapping at the origin, It is the total number of manifolds in this invention. That is, a manifold containing three components: H (hyperbolic), S (spherical), and E (Euclidean). (Lowercase) It is the index of the manifold, indicating which manifold the computation is performed on. is the spatial curvature of the current manifold, a learnable parameter whose core function is to define the geometric constraints of the exponential mapping, projecting a flat vector (MLP output) located in the tangent space at the origin onto a non-Euclidean manifold surface with specific geometric properties. (Symbol) The Cartesian product operator, representing the product manifold, is used to combine and stitch together the projection results of multiple independent manifold spaces into a unified feature representation of the product manifold. k For client indexing.

[0026] The above formula uniformly describes the process of mapping features to a "product manifold" composed of hyperbolic, spherical, and Euclidean spaces. In actual computational logic, this is manifested as multi-branch parallel computation, and finally, the token sequences that are independently mapped are concatenated and input into the large language model.

[0027] (2) Then, structural refinement based on Ollivier-Ricci curvature is performed, that is, Ollivier-Ricci curvature is used as a geometric prior to dynamically adjust the edge weights. In order to incorporate geometric priors into feature propagation, this invention calculates the Ollivier-Ricci curvature of the source image. Although the theoretical definition is based on the Wasserstein distance, in practice, this invention uses a combined approximation method based on triangle counting and node degree for calculation, taking the average of the lower and upper bounds of the combination of triangle counting and node degree as the estimated value of the Ollivier-Ricci curvature. in Representing nodes respectively and nodes The degree; Represents containing edges The number of triangles, i.e., nodes and The number of common neighbors; Operator , indicates that only the positive value is taken. Represents a node and nodes Between the edges The curvature value, represents the lower bound, represents the upper bound.

[0028] The curvature value is used to dynamically adjust the edge weight, so that the information propagation can perceive the local geometric characteristics. For the edge on the manifold , the curvature-adjusted is defined as: wherein, represents the potential influence strength of the neighbor node on the central node under the current geometric perspective, is the conversion of the basic edge weight to space as a reference score, which adaptively incorporates geometric curvature correction information through a neural network, is a manifold-specific neural network for learning how to convert curvature to weight correction value in this geometric space, and the normalized edge weight is calculated by the Softmax function: wherein, represents the potential influence strength of the neighbor node on the central node under the current geometric perspective, represents the neighbor node set of the central node .

[0029] (3) Finally, the geometric feature aggregation and mapping on the manifold are performed, that is, the node features are aggregated using the edge weight to realize geometric-aware propagation, and then mapped back to the Euclidean tangent space to obtain the semantic token sequence. Specifically, it includes: The updating process of node representation is completed in the tangent space of the Riemannian manifold with the central node as the reference. The neighbor node features are mapped to the tangent space of the central node features and then aggregated in the tangent space after weighting, and then the aggregation result is projected back to the manifold using the exponential mapping to realize geometric-aware propagation and updating, which is expressed as: wherein, represents the edge weight of the edge of the graph data of the kth client on the edge manifold , and is the current manifold spatial curvature, Indicates the features after aggregation. Indicates the update symbol. This represents the updated features of the central node; the process is executed iteratively twice by default, allowing the node embedding to fully integrate the geometric structure information of the neighborhood.

[0030] Nodes on each manifold This means first mapping back to the Euclidean space via the logarithm at the origin 0, and then using a manifold-specific encoder. (MLP) is transformed into a token vector with the same input dimension as the large language model: in, Indicates the first The node at the th manifold components The generated token vector. It is the final mapping product that maps and aligns the refined node representations in the non-Euclidean manifold space to the word embedding space of the large language model.

[0031] In this embodiment, local task learning based on semantic token sequences using a large language model is employed to update projector parameters and compute curvature importance vectors, including: To preserve the inductive bias of different geometric spaces, semantic token sequences from hyperbolic, spherical, and Euclidean spaces are used. Maintaining independence and linearly concatenating them in a preset order, these features are then incorporated into the Prompt template as geometrically enhanced features (i.e., replacing the features in the Prompt template). <graph>placeholder) and input into the parameter frozen large language model to obtain the predicted output of the local task.

[0032] The client computes the loss function based on the local task and updates the projector parameters. For node classification and link prediction tasks, the training objective is to align the LLM's prediction with the graph-based supervision signal. The invention uses the standard cross-entropy loss, which is defined as: where, is the length of the target sequence, is the output logit matrix of the LLM on the vocabulary , which has dimensions (position number vocabulary size). is the Logit value in the model's prediction at the th position corresponding to the correct answer Token . The denominator refers to the exponential sum of the model's scores for all possible Tokens in the vocabulary .

[0033] Curvature importance vector is obtained by calculating the gradient sensitivity of the projector parameters with respect to the geometric constraints, and is processed by the square term and normalized, where, represents the kth client, is the index of the manifold, , is the total number of projector parameters , and the curvature importance vector quantifies the importance contribution of each parameter dimension in capturing local geometric curvature features.

[0034] The server uses a shared subspace aggregation mechanism to realize projector parameter aggregation.

[0035] The server receives the updated parameters of the local projector uploaded by each client and the corresponding curvature importance vector . Due to the task heterogeneity of clients in federated learning, directly performing simple parameter averaging often leads to conflicts in the representation of geometric features. Therefore, the invention proposes a shared subspace aggregation mechanism, which independently aggregates the hyperbolic, spherical, and Euclidean three sets of manifold spaces. By using the curvature importance vector of each parameter dimension to fine-tune the update increment. Specifically, it includes: (1) Before performing curvature weighted whitening, a global parameter update matrix is constructed based on the projector parameters of all clients, including: adding the increments of the projector parameters uploaded by all clients Stacking, forming a global parameter update matrix , which represents the total amount of knowledge learned by all clients in the current round, the amount of parameter updates The definition is: where, represents the latest projector parameter obtained by the th client after completing the local task training in the th round, represents the global projector parameter sent by the server in the th round, is a vectorization operator, which is used to expand the matrix format parameter into a one-dimensional column vector for subsequent weighted aggregation calculation.

[0036] (2) Curvature weighted whitening, that is, calculating the curvature importance vector weighted whitening update matrix, which specifically includes: In order to accurately extract shared geometric features and suppress statistical noise in the aggregation process, first use the global parameter update matrix to calculate the empirical variance matrix of each parameter dimension to measure the fluctuation degree of different parameter dimensions: Then, combined with the average curvature importance vector of each client , construct a geometric perception scaling matrix This matrix adjusts the parameter weight by comprehensively considering the statistical stability and geometric sensitivity of the dimension: Introduce the re-calibration weight to balance the uneven contribution of each client due to the difference in update strength, the size of the re-calibration weight is proportional to the update strength, based on the re-calibration weight and the geometric perception scaling matrix , construct the whitening update matrix : where, represents function, used to construct a diagonal matrix, superscript T represents matrix transpose, is a very small positive number, its main role is to prevent the denominator from being zero, so as to ensure the stability in numerical calculation, meet constraint condition, is a weight vector composed of the re-calibration weights of all clients .

[0037] (3) Two-stage singular value decomposition to extract subspace, that is, based on the whitening update matrix, two-stage SVD decomposition is performed on the projector parameters of all clients to extract the shared geometric subspace.

[0038] The total number of parameters caused by the multi-layer neural network of the projector To address the problems of excessively high SVD computational overhead, this invention adopts a block aggregation strategy, which increments parameters based on the physical network structure. Map back to hierarchical parameter block ( (as a layer index), and reconstruct it into the original weight matrix form. This mechanism not only significantly reduces computational complexity and memory consumption, but also allows for fine-tuning of curvature-weighted whitening and two-stage SVD for statistical properties at different levels, thereby improving the robustness of feature extraction.

[0039] To extract consistent geometric features across tasks, this invention performs a two-stage singular value decomposition for each parameter block. During this process, a preset rank-preserving ratio is used. Dynamically determine the rank to be retained at each level ,in The projector Number of output neurons in the layer weight matrix: Local feature extraction: for each client Specific parameter blocks Whitening update matrix Perform SVD decomposition, retaining the previous... The left singular vectors are used as the local basis matrix. This step captures the dominant line space orientation of a single client in the whitening space; Global shared extraction: Concatenate the local bases of all clients into a matrix. ,right Perform a second SVD decomposition and retain the previous values. The left singular vectors serve as the basis matrix of the shared geometric subspace. .

[0040] Through these two stages of SVD decomposition, this invention can accurately identify and extract geometrically consistent shared directions from heterogeneous client updates. Based on this, performing a global model update can effectively filter local task noise and achieve more accurate alignment between multidimensional geometric priors and the general semantic space.

[0041] (4) Consistent aggregation and global parameter update, that is, the global update is achieved by projecting the projector parameters into the shared geometric subspace and then weighting the aggregation.

[0042] After extracting the shared subspace basis matrix of manifold components ,the server no longer averages directly in the original parameter space, but aggregates within the consistency subspace to completely filter out the noise interference caused by heterogeneous tasks of clients.

[0043] Diagonal projection operator is a geometric constraint mechanism built on the whitened space. It aims to solve the gradient direction conflict caused by data heterogeneity by forcing the client updates to be mapped to the shared subspace. This diagonal projection operator is directly calculated based on the shared geometric subspace basis matrix extracted by the server through two-stage SVD , which only retains the manifold structure direction with global consensus, ensuring the consistency and stability of model updates in a heterogeneous environment.

[0044] Since the diagonal projection operator is only effective in the whitened space where the parameter magnitude has been normalized, its direct application in the original space will cause direction distortion due to uneven variance distribution. Therefore, based on the diagonal projection operator and the geometric-aware scaling matrix , the projector parameters are executed in a whitening-projection-inverse whitening structure, i.e., by mapping the update to the whitened space, applying the projection operator to filter noise, and finally using the inverse whitening transformation to restore the original magnitude of the parameters: The server uses the local sample weights of each client ( and represent the vertex set in the local graph data owned by the th and th client, respectively) to perform a weighted average of the projected parameters to calculate the final global parameter update : The server uses this global parameter update to update the global projector parameters, and the updated global projector parameters are distributed to each client for use in the next round of federated learning.

[0045] where is the total number of clients, is the weighted parameter update, and These are the global projector parameters for rounds t+1 and t, respectively.

[0046] Through the above technical solutions, this invention solves the problem of structural understanding under data scarcity by using multi-manifold geometric projection on the client side, and solves the problem of parameter conflict under heterogeneous tasks by using shared subspace aggregation on the server side, thus realizing efficient and robust federated graph learning.

[0047] The embodiment also provides an enhanced federated graph learning method for large language models targeting heterogeneous tasks, which employs the aforementioned system, such as... Figure 2 As shown, it includes the following steps: S1, the client projects local private graph data into a semantic token sequence through a projector, uses a large language model to perform local task learning based on the semantic token sequence to update the projector parameters and calculate the curvature importance vector, and sends the projector parameters and curvature importance vector to the server. S2, the server uses a shared subspace aggregation mechanism to aggregate projection parameters; the globally updated projector parameters are then distributed to each client for use in the next round of federated learning.

[0048] The methods and systems provided in the above embodiments effectively solve the problem of deep representation of complex graph topologies under limited sample conditions by constructing a product manifold containing hyperbolic, spherical, and Euclidean spaces and dynamically adjusting edge weights using Ollivier-Ricci curvature. This allows for more accurate reconstruction of hierarchical and cyclic structures in the data, avoiding semantic distortions of complex graph structures.

[0049] The method and system provided in the above embodiments are the first to use a shared subspace extraction mechanism based on two-stage SVD, which can separate geometrically consistent features from heterogeneous parameters, thereby effectively overcoming gradient conflicts between differentiated tasks such as node classification and link prediction, and avoiding negative migration.

[0050] The methods and systems provided in the above embodiments propose a parameter whitening strategy based on curvature importance. By standardizing and weighting the parameters according to their sensitivity to geometric structures, accurate alignment of model parameter semantics is achieved under heterogeneous tasks.

[0051] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.< / graph>

Claims

1. A large language model enhanced federated graph learning system for heterogeneous tasks, characterized in that, It includes a server and multiple clients holding local private graph data; The client is used to project local private graph data into a semantic token sequence through a projector. Specifically, this includes: constructing a product manifold containing hyperbolic, spherical, and Euclidean spaces, calculating the node features of the graph data projected onto the product manifold, dynamically adjusting edge weights using Ollivier-Ricci curvature as a geometric prior, using the edge weights to perform geometric aggregation on the product manifold to achieve geometric perception propagation, and then mapping it back to Euclidean tangent space to obtain a semantic token sequence; using a large language model to perform local task learning based on the semantic token sequence to update the projector parameters and calculate the curvature importance vector, and sending the projector parameters and curvature importance vector to the server. The server-side implementation uses a shared subspace aggregation mechanism to aggregate projector parameters. Specifically, it includes: constructing a global parameter update matrix based on the projector parameters of all clients, calculating the parameter update matrix after curvature importance vector weighted whitening, performing two-stage SVD decomposition on the projector parameters of all clients based on the whitened parameter update matrix to extract the shared geometric subspace, projecting the updated projector parameters of each client using the extracted shared geometric subspace, and then obtaining the global update amount through weighted aggregation to achieve global parameter update; and distributing the updated global projector parameters to each client for the next round of federated learning.

2. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, Constructing a product manifold encompassing hyperbolic, spherical, and Euclidean spaces and calculating the nodal features of the graph data projected onto the product manifold, including: Defining product manifolds For hyperbolic space , spherical space and Euclidean space : ； wherein, represents a hyperbolic curvature, used to capture hierarchical structures; represents a spherical curvature, used to capture cyclic structures; is the intrinsic dimensionality each independent geometric space component has when building the product manifold; By manifold-specific encoders Mapping original node features in graph data to tangent spaces of each manifold, and then projecting them onto the manifolds using exponential maps to get node features projected onto the product manifold Mapping original node features in graph data to tangent spaces of each manifold, and then projecting them onto the manifolds using exponential maps to get node features projected onto the product manifold , which can be expressed as: ； wherein, denotes the exponential map at the origin, is the total number of manifolds, and lowercase is the index of the manifold, representing the number of the manifold being computed, is the spatial curvature of the current manifold, is a learnable parameter, and the symbol represents the Cartesian product operator of the product manifold, used to combine the projection results of multiple independent manifold spaces into a unified product manifold feature representation, k is the client index.

3. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, Using Ollivier-Ricci curvature as a geometric prior to dynamically adjust edge weights includes: The curvature value is calculated using a combined approximation method based on triangle counting and node degree. The average of the lower and upper bounds of the combination of triangle counting and node degree is taken as the estimated value of the Ollivier-Ricci curvature value. ； wherein, denotes a node and a node between which an edge has a curvature value, denotes a lower bound, denotes an upper bound; For edges on a manifold whose curvature is adjusted to is defined as​ ； where, represents the potential influence strength of a neighbor node on the central node at the current geometric view, is the base edge weight converted to space as a baseline score, and then the model adaptively incorporates the geometric curvature correction information through a neural network, is a manifold-specific neural network that learns how to convert the curvature into a weight correction value in this geometric space, and then the normalized edge weight is calculated through a Softmax function ； wherein, represents the potential influence strength of a neighbor node on the center node at the current geometric perspective, represents the set of neighbor nodes of the center node.

4. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, Geometric aggregation of node features on the product manifold using edge weights is performed to achieve geometric perceptual propagation, which is then mapped back to the Euclidean tangent space to obtain a semantic token sequence, including: First, the neighbor node features are mapped to the tangent space of the center node features using a log map and weighted aggregated in the tangent space. Then, the aggregated result is projected back to the manifold using an exponential map, which realizes the geometric perception propagation and update, and is formulated as ​ ； ； wherein, denotes the edge weight of the kth client's graph data on the edge manifold upper edge , is the spatial curvature of the current manifold , denotes the aggregated feature, denotes the update symbol, denotes the updated center node feature; Nodes on each manifold Represents first mapping back to the Euclidean tangent space at the origin 0 via a logarithmic map, then through manifold-specific encoders Transformed into Token vectors consistent with the input dimension of the large language model: ； wherein, represents the Token vector generated by the th node in the th manifold component .

5. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, The projector parameters and curvature importance vectors are updated using local task learning based on semantic token sequences using a large language model, including: Semantic token sequences from hyperbolic, spherical, and Euclidean spaces are kept independent and linearly concatenated in a preset order. These sequences are then used as geometrically enhanced features and incorporated into the Prompt template. Finally, they are input into a large language model with frozen parameters to obtain the prediction output for the local task. Curvature importance vector By computing the gradient sensitivity of the projector parameters with respect to the geometric constraints, and processing with square terms and normalization, where, represents the kth client, is the index of the manifold, , is the parameter of the projector The total amount of parameters, the curvature importance vector quantifies the importance contribution of each parameter dimension in capturing the local geometric curvature features.

6. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, A global parameter update matrix is ​​constructed based on the projector parameters of all clients, including: The increment of all the projector parameters uploaded by all clients Stacking, forming a global parameter update matrix , which represents the total amount of knowledge learned by all clients in the current round, the parameter update amount The definition is: ； wherein, denotes the th client's latest projector parameter obtained after the local task training in the th round, denotes the global projector parameter sent by the server in the th round, is a vectorization operator, which serves to expand the matrix-formatted parameter into a one-dimensional column vector.

7. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, Calculate the parameter update matrix after weighted whitening of the curvature importance vector, including: First, the global parameter update matrix is calculated The empirical variance matrix of each parameter dimension is calculated To measure the volatility of different parameter dimensions: ； Subsequently combine the averaged curvature importance vectors for each client , construct a geometry-aware scaling matrix : ； Introducing recalibration weights To balance the contribution of each client due to the difference in update intensity The size of the recalibration weight is proportional to the update intensity And the geometric perception scaling matrix Construct a whitening update matrix : ； in, express The function is used to construct a diagonal matrix, where the superscript T indicates matrix transpose. This represents a positive number used to prevent the denominator from being zero, satisfying... The constraints, The recalibrated weights of all clients The weight vector is composed of these components.

8. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, Based on the whitened parameter update matrix, a two-stage SVD decomposition is performed on the projector parameters of all clients to extract the shared geometric subspace, including: Local feature extraction: Perform SVD decomposition on the whitened update matrix of each client's specific parameter block, retaining the first K left singular vectors as the local basis matrix ;​​​​ Global shared extraction: concatenating all clients' local bases into a matrix , performing a second SVD decomposition and retaining the first left singular vectors as shared geometric subspace basis matrix .​ 9. The heterogeneous task-oriented large language model enhanced federated graph learning system of claim 1, wherein, The extracted shared geometric subspace is used to project the projector parameter update values ​​of each client, followed by weighted aggregation to obtain the global update value and implement global parameter updates, including: Sharing geometric subspace basis matrix Computing a diagonal projection operator ; Based on a diagonal projection operator and a geometrically aware scaling matrix to the projector parameters Performing a consistent projection with a whitening-projection-inverse whitening structure to obtain the projected parameters : ； Utilizing local sample weights of each client On the projected parameters Weighted aggregation to achieve global update: ； ； wherein, is the total number of clients, is the weighted parameter update amount, and are the global projector parameters of the t+1th round and the tth round, respectively.

10. A large language model enhanced federated graph learning method for heterogeneous tasks, characterized in that, The method employs the system described in any one of claims 1-9 and includes the following steps: The client projects local private graph data into a semantic token sequence through a projector, specifically including: constructing an accumulated manifold containing hyperbolic, spherical and Euclidean spaces and calculating node features of the graph data projected onto the accumulated manifold, taking Ollivier-Ricci curvature as a geometric prior to dynamically adjust edge weights, and using the edge weights to perform geometric aggregation on the node features on the accumulated manifold to realize geometric perception propagation and then map back to the Euclidean tangent space to obtain a semantic token sequence; updating the projector parameters and calculating the curvature importance vector based on the semantic token sequence through local task learning of a large language model, and sending the projector parameters and the curvature importance vector to the server; The server uses a shared subspace aggregation mechanism to realize projector parameter aggregation, specifically including: constructing a global parameter update matrix based on the projector parameters of all clients, and calculating the parameter update matrix after weighted whitening of the curvature importance vector, performing two-stage SVD decomposition on the projector parameters of all clients based on the whitened parameter update matrix to extract a shared geometric subspace, and projecting the projector parameter update amount of each client using the extracted shared geometric subspace, then obtaining a global update amount through weighted aggregation and realizing global parameter update; the globally updated projector parameters are distributed to each client for use in the next round of federated learning.

Images