Graph federated learning method, system and storage medium based on validation set accuracy optimization

Abstract

This invention discloses a graph federated learning method, system, and storage medium based on validation set accuracy optimization. The method includes the following steps: generating a basic training model using VGAE; constraining local models to be closer to the global model and farther away from the previous local model based on a model similarity assessment method; calculating new local model parameters and fine-tuning them based on the local validation set; and finally, a server collecting local model parameters from each client for global aggregation and update. This invention addresses the problems of the prior art in lacking targeted optimization for non-Euclidean data and resulting in reduced model performance.

Description

Technical Field

[0001] The present invention belongs to the field of federated learning and privacy protection technology, and in particular relates to a graph federated learning method, system, and storage medium based on verification set accuracy optimization. Background Art

[0002] Federated Learning (FL) is an emerging distributed machine learning paradigm that allows clients to collaboratively train shared global or personalized models without sharing their original training data. In FL, each client uploads only its own model parameters or model gradients to the server in each round. The server then averages and aggregates the data to update the local model. No original data is exchanged during this process.

[0003] New approaches to graph neural networks (GNNs) have been used to process graph data, achieving state-of-the-art performance in many graph-based tasks, such as link prediction, node classification, community detection, and graph classification. However, due to privacy and large-scale data collection issues, recent research has leveraged federated learning on graph-structured data to share latent knowledge and address the knowledge scarcity problem. This is known as graph federated learning (FGL), such as Federated ScopeGNN and FedSage.

[0004] However, the aforementioned graph federated learning methods do not consider the statistical heterogeneity and personalization of the global model when processing non-Euclidean data, and ignore the inherent structural heterogeneity within the model, data, and devices, which leads to a decline in the overall performance of the model and worsens the performance of the global FL model on a single client.

[0005] In order to improve the generalization ability of non-Euclidean data (such as graph data) and enhance model performance, a graph federated learning method, system and storage medium based on validation set accuracy optimization are proposed. Summary of the Invention

[0006] The embodiments of the present invention propose a graph federated learning method, system, and storage medium based on validation set accuracy optimization to at least solve the problems of lack of targeted optimization for non-Euclidean data and degradation of model performance in related technologies.

[0007] According to one embodiment of the present invention, a graph federated learning method based on validation set accuracy optimization is proposed, including:

[0008] Use VGAE to generate a basic training model;

[0009] Constrain the local model to be close to the global model and away from the local model of the previous round according to the model similarity evaluation method;

[0010] Calculate new local model parameters and fine-tune them based on the local validation set;

[0011] The server collects the local model parameters of each client for global aggregation and update.

[0012] In an exemplary embodiment, the generating of the basic training model using VGAE comprises the steps of:

[0013] Select VGAC as the base model and construct a three-layer graph convolutional network encoder;

[0014] Gaussian distribution is used to calculate the probability distribution of the node's potential representation;

[0015] Train itself by optimizing the variational lower bound;

[0016] The loss functions for the link prediction task and the node classification task are calculated separately to form the basic training model.

[0017] In an exemplary embodiment, constraining the local model to be close to the global model and away from the local model of the previous round according to the model similarity evaluation method includes the steps of:

[0018] Calculate the similarity function between the local model and the global model according to the model similarity evaluation method;

[0019] Calculate the similarity function between the local model of this round and the model of the previous round according to the model similarity evaluation method;

[0020] Calculate the model comparison loss function based on the similarity function between the local model of this round and the global model, and the similarity function between the local model of this round and the local model of the previous round;

[0021] Calculate the local model constraint loss function based on the model contrast loss function;

[0022] The local model is trained according to the local model constraint loss function.

[0023] In an exemplary embodiment, the model similarity evaluation method comprises the steps of:

[0024] Calculate the cosine similarity based on the cosine of the angle between the model vectors;

[0025] Calculate the Euclidean distance based on the distance between model vectors;

[0026] The neighborhood overlap ratio is calculated based on the overlap ratio of neighborhoods between models;

[0027] Calculate the Pearson correlation coefficient based on the linear correlation of the model vectors;

[0028] The similarity function between the models is calculated based on cosine similarity and / or Euclidean distance and / or neighborhood overlap rate and / or Pearson correlation coefficient.

[0029] In an exemplary embodiment, the calculating of new local model parameters and fine-tuning based on a local validation set comprises the steps of:

[0030] The local model parameters and the global model parameters are superimposed in proportion to obtain new local model parameters after local aggregation;

[0031] Verify the new local model parameters after local aggregation based on the local validation set data;

[0032] When the preset adjustment conditions are met, the scale of the global model is adjusted and the new local model parameters after local aggregation are recalculated.

[0033] In an exemplary embodiment, the verification of the new local model parameters after local aggregation is to sample a batch of data at the beginning and verify the model with the sampled data during the model training process; the preset adjustment condition is lower than the preset model test accuracy and / or lower than the preset model convergence speed.

[0034] In an exemplary embodiment, the server collects local model parameters of each client for global aggregation and update, including the steps of:

[0035] The server collects the trained local model parameters of each client;

[0036] Calculate the change weight value according to the change of each local model parameter;

[0037] Integrate the parameters of each local model according to the change weight value to form a new global model;

[0038] The global model is sent to each client and the next round of training continues until the model converges.

[0039] In an exemplary embodiment, the step of calculating the change weight value according to the change of each local model parameter comprises the following steps:

[0040] Calculating a parameter change rate based on a parameter change amount and / or a parameter change ratio between the local model parameters and the local model parameters of the previous round;

[0041] Calculate the parameter change acceleration rate according to the parameter change acceleration of the local model parameter, the local model parameter of the previous round, and the local model parameters of the previous two rounds;

[0042] The change weight value of the local model is calculated according to the parameter change rate and / or the parameter change acceleration rate.

[0043] A computer-readable storage medium stores a computer program for electronic data exchange, wherein the computer program causes a computer to execute the above method.

[0044] According to another embodiment of the present invention, a graph federated learning system based on validation set accuracy optimization is provided, including:

[0045] processor;

[0046] Memory;

[0047] as well as

[0048] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, the programs causing the computer to perform the above method.

[0049] The graph federated learning method, system, and storage medium based on validation set accuracy optimization of the present invention have the following advantages:

[0050] (1) The contrastive learning method in self-supervised learning is used to compare the global model and the historical local model. Compared with the traditional local model training method, the degree of overfitting of the GNNs model is reduced and the performance of the GNNs model in the federated scenario is improved.

[0051] (2) The similarity function between models is calculated based on cosine similarity and / or Euclidean distance and / or neighborhood overlap rate and / or Pearson correlation coefficient. Compared with the traditional technical solution of evaluating model similarity only based on pre-similarity, the accuracy of model comparison training can be effectively improved.

[0052] (3) Compared with traditional local model optimization methods, model aggregation and optimization based on local validation set data can effectively solve the phenomenon of model dispersion and aggregation, so that the entire global model can be further optimized towards the global optimal goal, thereby improving the performance of the global model.

[0053] (4) The decoders of each GNNs are retained locally and used as the personalized layer of the model. The change weight values ​​of each local model are calculated according to the parameter change rate and / or parameter change acceleration rate of the local model and the global model is updated accordingly. Compared with the traditional federated learning method, the personalized ability of the local model is improved while retaining the global knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flowchart of a graph federated learning method based on validation set accuracy optimization according to an embodiment of the present invention;

[0055] Figure 2 is a flowchart of sub-step S01 of an embodiment of the present invention;

[0056] Figure 3 is a flowchart of sub-step S02 of an embodiment of the present invention;

[0057] Figure 4 is a flow chart of a model similarity evaluation method according to an embodiment of the present invention;

[0058] Figure 5 is a flowchart of sub-step S03 of an embodiment of the present invention;

[0059] Figure 6 is a flowchart of sub-step S04 of an embodiment of the present invention;

[0060] Figure 7 is a flowchart of sub-step S042 of an embodiment of the present invention;

[0061] Figure 8 This is a schematic diagram of the structure of a graph federated learning system based on validation set accuracy optimization in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the invention, but are not intended to limit the present invention in any form. It should be noted that those skilled in the art may make several changes and modifications without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0063] A graph federated learning method based on validation set accuracy optimization according to an embodiment of the present invention is shown in the flowchart Figure 1 As shown, the steps include:

[0064] Step S01: Generate a basic training model using VGAE;

[0065] Step S02: constraining the local model to be close to the global model and away from the local model of the previous round according to the model similarity evaluation method;

[0066] Step S03: Calculate new local model parameters and fine-tune them based on the local validation set;

[0067] Step S04: The server collects local model parameters of each client and performs global aggregation and update.

[0068] In an exemplary embodiment, the step S01 is as shown in the flow chart. Figure 2 Shown, including:

[0069] Step S011: select VGAC as the basic model and construct a three-layer graph convolutional network encoder;

[0070] Step S012: Calculate the probability distribution of the node's potential representation using Gaussian distribution;

[0071] Step S013: training itself by optimizing the variational lower bound;

[0072] Step S014: Calculate loss functions for the link prediction task and the node classification task respectively to form a basic training model.

[0073] In this embodiment, real datasets usually exhibit various characteristics, including different types and fields, and contain a large amount of data. To meet this challenge, VGAC can reconstruct and generate new graph data to achieve generalization and functionality in different fields. In addition, it utilizes a variational inference method that effectively processes large graph datasets. Therefore, VGAC is selected as the base model. The encoder is constructed as a direct three-layer graph convolutional network (GCN) to encode the topological structure a and node features X. Finding the probability distribution of the potential representation of a node is often a challenge. Therefore, the distribution needs to be approximated. Gaussian distribution, as the most common distribution, is often used to solve this problem.

[0074]

[0075]

[0076] where μ i and σ i Denote the l-dimensional mean and standard deviation vectors corresponding to node u. The distribution parameters here can be learned from the two-layer GCN as shown below

[0077] μ=GCN μ (GCN1(A,X)),A)

[0078] log σ =GCN σ (GCN1(A,X)),A)

[0079] Z n*d =reparametrize(μ,σ)

[0080] Among them, GCN μ and GCN σ are different layers to generate μ u and σ u The matrix of GCN1 is a common layer for generating the lower embedding of nodes. Z is the node representation sampled from the distribution network, reparametrize(μ,σ)=μ+∈*σ,∈~N(0,1). Given Z, the decoder dcc(,) in VGAC is a simple inner product as follows

[0081]

[0082] in is the reconstructed adjacency matrix, z v is the embedding of node v. VGAC trains itself by optimizing the variational lower bound (CLBO):

[0083] L ELBO =E q(Z|γ) [p(A|Z)]-KL[p(Z|X,A)||(Z)]

[0084] Among them, for the link prediction task, its loss function L o1 for:

[0085]

[0086] Where n represents the number of nodes. In addition, a two-layer MLP is used as the readout function for node classification. The classification loss function is L o2 Expressed as:

[0087]

[0088] where n k represents the probability that the node belongs to the kth class, is the prediction result. The loss function of the task is given by L o Unified representation, thus obtaining the basic training model. The loss function is summarized as:

[0089]

[0090] In an exemplary embodiment, the step S02 is as shown in the flowchart. Figure 3 Shown, including:

[0091] Step S021: Calculate the similarity function between the local model and the global model in this round according to the model similarity evaluation method;

[0092] Step S022: Calculate the similarity function between the local model of this round and the model of the previous round according to the model similarity evaluation method;

[0093] Step S023: Calculate the model comparison loss function based on the similarity function between the local model of this round and the global model and the similarity function between the local model of this round and the local model of the previous round;

[0094] Step S024: Calculate the local model constraint loss function according to the model comparison loss function;

[0095] Step S025: Perform local model training according to the local model constraint loss function.

[0096] In this embodiment, we use the MOON comparison method in the field of computer vision, that is, based on the intuition that the global model has better generalization performance than the local model, we approximate the local model and the global model and move away from the local historical model to perform comparative learning. i represents the local model parameters, Θ represents the global model parameters, z, z g 、z p The nodes correspond to the current round of the global model through the encoder, the local model of the current round, and the local model of the previous round. Therefore, the model contrast loss can be displayed as follows.

[0097]

[0098] Where sim(·) is the similarity function obtained by the model similarity evaluation method, and τ is the temperature parameter. Generally speaking, the loss function is defined as follows:

[0099] L=L o +λL con

[0100] Where λ is a hyperparameter set empirically, L o Depends on the task type.

[0101] In this embodiment, model comparison loss and training loss are jointly optimized. After the first round of training, the trained model is retained and stored in a queue. A new model is added with each training round. If the upper limit of the number of comparison models is reached, the oldest entry model is polled. In each training round, the global model obtained from the server and the locally saved model of the current round are compared with the currently trained model for learning.

[0102] In an exemplary embodiment, the sim(·) function is obtained according to the model similarity evaluation method, and the flow chart is as follows: Figure 4 Shown, including:

[0103] S10, calculating the cosine similarity based on the cosine value of the angle between the model vectors;

[0104] S11, calculating the Euclidean distance based on the distance between the model vectors;

[0105] S12, calculating the neighborhood overlap rate according to the overlap ratio of the neighborhoods between the models;

[0106] S13, calculating the Pearson correlation coefficient based on the linear correlation of the model vector;

[0107] S14. Calculate a similarity function between the models based on cosine similarity and / or Euclidean distance and / or neighborhood overlap rate and / or Pearson correlation coefficient.

[0108] In this embodiment, the cosine similarity is calculated based on the cosine value of the angle between the model vectors, which is the cosine similarity calculated based on the ratio of the vector product and the module length product between the two models. The cosine similarity is represented by c.

[0109] The Euclidean distance is calculated based on the distance between the model vectors, and the Euclidean distance is represented by the variable d.

[0110] The neighborhood overlap rate is calculated based on the overlap ratio of neighborhoods between models, and the neighborhood overlap rate is represented by the variable n.

[0111] The calculation of the Pearson correlation coefficient based on the linear correlation of the model vectors is to obtain the Pearson correlation coefficient between the two model vectors according to the Pearson correlation coefficient calculation formula, and the Pearson correlation coefficient is represented by p.

[0112] The similarity function between models calculated based on cosine similarity and / or Euclidean distance and / or neighborhood overlap rate and / or Pearson correlation coefficient is a similarity function between models calculated based on the positive correlation between cosine similarity and / or Euclidean distance and / or neighborhood overlap rate and / or Pearson correlation coefficient and similarity between models, and the similarity function between models is represented by sim.

[0113] A1 to A15 in Table A represent different implementations of calculating the similarity function between models. For ease of expression, the cosine similarity c, Euclidean distance d, neighborhood overlap rate n, and Pearson correlation coefficient p are calculated using the method described in any of the above implementations.

[0114] Table A Different implementations of similarity functions between models

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] In an exemplary embodiment, the step S03 is as shown in the flowchart. Figure 5 Shown, including:

[0122] Step S031: Superimpose the local model parameters and the global model parameters in proportion to obtain new local model parameters after local aggregation;

[0123] Step S032: verify the new local model parameters after local aggregation based on the local verification set data;

[0124] Step S033: When the preset adjustment conditions are met, the scale of the global model is adjusted and the new local model parameters after local aggregation are recalculated.

[0125] In an exemplary embodiment, the verification of the new local model parameters after local aggregation is to sample a batch of data at the beginning and verify the model with the sampled data during the model training process; the preset adjustment condition is lower than the preset model test accuracy and / or lower than the preset model convergence speed.

[0126] In this embodiment, in the previous federated learning process, all clients directly updated their local parameters after receiving the gradients and model parameters from the server. To some extent, the quality of the local model deteriorated a little instead of benefiting from the global model. To cope with this situation, fine-tuning will be performed during the aggregation process to increase the gains and reduce the losses, so that each client model can achieve better performance. Formally,

[0127]

[0128] in is the t-round model parameter on the i-th client, is the new local parameter after local aggregation, is a hyperparameter used for learning.

[0129] For client i, at the beginning, a batch of data D i During the model training process, the local model is first updated with the parameters obtained from the server, and then the sampled data D is used to i Verify the model. If the model accuracy is lower than the preset model test accuracy and / or lower than the preset model convergence speed (i.e., the preset adjustment conditions are met), the local model will be re-aggregated, the scale of the global model will be adjusted, and the parameters of the new local model after local aggregation will be recalculated.

[0130] In an exemplary embodiment, the step S04 is as shown in the flowchart. Figure 6 Shown, including:

[0131] Step S041: The server collects the trained local model parameters of each client;

[0132] Step S042: Calculate the change weight value according to the change of each local model parameter;

[0133] Step S043: Integrate the local model parameters according to the change weight value to form a new global model;

[0134] Step S044: Send the global model to each client and continue the next round of training until the model converges.

[0135] In this embodiment, a trusted central server is assumed. Initially, the global model sends an initially generated model to each client. After client training is complete, the server collects the trained local model parameters of each client, calculates change weights based on the changes in each local model parameter, and then performs a weighted average of the local model parameters based on the change weights. This is then combined to form a new global model, which is then sent to each client and training continues until the model converges.

[0136] In an exemplary embodiment, the sub-step S042 is as shown in the flowchart. Figure 7 Shown, including:

[0137] Step S0421: Calculate the parameter change rate based on the parameter change amount and / or parameter change ratio of the local model parameters compared with the local model parameters of the previous round;

[0138] Step S0422: Calculate the parameter change acceleration rate based on the change in the local model parameter change value between adjacent rounds;

[0139] Step S0423: Calculate the change weight value of the local model according to the parameter change rate and / or the parameter change acceleration rate.

[0140] In this embodiment, the parameter change rate is calculated based on the parameter change amount and / or parameter change ratio of the local model parameters and the local model parameters of the previous round. The parameter change value is calculated based on the ratio of the parameter change amount of the local model parameters and the local model parameters of the previous round to the preset parameter threshold and the positive correlation between the parameter change rate, or the parameter change rate is calculated based on the positive correlation between the parameter change ratio of the local model parameters and the local model parameters of the previous round and the parameter change rate, or the parameter change rate is calculated based on the ratio of the parameter change amount of the local model parameters and the local model parameters of the previous round to the preset parameter threshold and the positive correlation between the parameter change ratio and the parameter change rate.

[0141] The method of calculating the parameter change acceleration rate based on the change in the local model parameter change value of adjacent rounds is to record the change in the local model parameter of the current round and the previous round in each round of training, calculate the average change in multiple groups of parameter change values ​​within a preset time period (that is, the change in the local model parameter change value of adjacent rounds), and calculate the parameter change acceleration rate based on the positive correlation between the average change in multiple groups of parameter change values ​​within the preset time period and the parameter change acceleration rate. For example, the local model parameter change value in the first and second rounds of training is a1, the local model parameter change value in the second and third rounds is a2, the local model parameter change value in the third and fourth rounds is a3, and the local model parameter change value in the fourth and fifth rounds is a4. Assuming that the preset time period includes the first to fifth rounds of training, the change in the local model parameter change value of adjacent rounds is calculated as a=(|a2-a1|+|a3-a2|+|a4-a3|) / 3. The parameter change acceleration rate is calculated based on the positive correlation between the change in the local model parameter change value a of adjacent rounds and the parameter change acceleration rate.

[0142] The method of calculating the change weight value of the local model based on the parameter change rate and / or the parameter change acceleration rate is any one of calculating the change weight value of the local model based on the positive correlation between the parameter change rate and the change weight value of the local model, calculating the change weight value of the local model based on the positive correlation between the parameter change acceleration rate and the change weight value of the local model, and calculating the change weight value of the local model based on the positive correlation between the parameter change rate and the parameter change acceleration rate and the change weight value of the local model.

[0143] A computer-readable storage medium according to an embodiment of the present invention stores a computer program for electronic data exchange, wherein the computer program enables a computer to execute the above method.

[0144] A graph federated learning system based on validation set accuracy optimization according to an embodiment of the present invention is shown in the structural diagram below: Figure 8 Shown, including:

[0145] processor;

[0146] Memory;

[0147] as well as

[0148] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, the programs causing the computer to perform the above method.

[0149] Of course, those skilled in the art should realize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. As long as they are within the scope of the present invention, any changes or modifications to the above embodiments will fall within the scope of protection of the present invention.

Claims

1. A graph federated learning method based on validation set accuracy optimization, characterized by: include: Use VGAE to generate a basic training model; Constrain the local model to be close to the global model and away from the local model of the previous round according to the model similarity evaluation method; Calculating new local model parameters and fine-tuning them according to a local validation set; calculating new local model parameters and fine-tuning them according to a local validation set, comprising the steps of: superimposing local model parameters and global model parameters in proportion to obtain locally aggregated new local model parameters; verifying the locally aggregated new local model parameters based on local validation set data; and adjusting the proportion of the global model and recalculating the locally aggregated new local model parameters when a preset adjustment condition is met; The verification of the new local model parameters after local aggregation is to sample a batch of data at the beginning and verify the model with the sampled data during the model training process; the preset adjustment condition is lower than the preset model test accuracy and / or lower than the preset model convergence speed; The server collects the local model parameters of each client for global aggregation and update.

2. The graph federated learning method based on validation set accuracy optimization according to claim 1 is characterized in that The method of generating a basic training model using VGAE includes the following steps: Select VGAE as the base model and construct a three-layer graph convolutional network encoder; Gaussian distribution is used to calculate the probability distribution of the node's potential representation; Train itself by optimizing the variational lower bound; The loss functions for the link prediction task and the node classification task are calculated separately to form the basic training model.

3. The graph federated learning method based on validation set accuracy optimization according to claim 1 is characterized in that The method of constraining the local model to be close to the global model and away from the local model of the previous round according to the model similarity evaluation method includes the following steps: Calculate the similarity function between the local model and the global model according to the model similarity evaluation method; Calculate the similarity function between the local model of this round and the model of the previous round according to the model similarity evaluation method; Calculate the model comparison loss function based on the similarity function between the local model of this round and the global model, and the similarity function between the local model of this round and the local model of the previous round; Calculate the local model constraint loss function based on the model contrast loss function; The local model is trained according to the local model constraint loss function.

4. The graph federated learning method based on validation set accuracy optimization according to claim 3 is characterized in that The model similarity evaluation method comprises the steps of: Calculate the cosine similarity based on the cosine of the angle between the model vectors; Calculate the Euclidean distance based on the distance between model vectors; The neighborhood overlap ratio is calculated based on the overlap ratio of neighborhoods between models; Calculate the Pearson correlation coefficient based on the linear correlation of the model vectors; The similarity function between the models is calculated based on cosine similarity and / or Euclidean distance and / or neighborhood overlap rate and / or Pearson correlation coefficient.

5. The graph federated learning method based on validation set accuracy optimization according to claim 1 is characterized in that The server collects local model parameters of each client for global aggregation and update, including the following steps: The server collects the trained local model parameters of each client; Calculate the change weight value according to the change of each local model parameter; Integrate the parameters of each local model according to the change weight value to form a new global model; The global model is sent to each client and the next round of training continues until the model converges.

6. The graph federated learning method based on validation set accuracy optimization according to claim 5 is characterized in that: The method of calculating the change weight value according to the change of each local model parameter comprises the steps of: Calculating a parameter change rate based on a parameter change amount and / or a parameter change ratio between the local model parameters and the local model parameters of the previous round; Calculate the parameter change acceleration rate based on the change in the local model parameter value between adjacent rounds; The change weight value of the local model is calculated according to the parameter change rate and / or the parameter change acceleration rate.

7. A computer-readable storage medium storing a computer program for electronic data exchange, wherein: The computer program enables a computer to execute the method according to any one of claims 1 to 6.

8. A graph federated learning system based on validation set accuracy optimization, characterized by include: processor; Memory; as well as One or more programs, wherein the one or more programs are stored in a memory and configured to be executed by the processor, the programs causing the computer to perform the method according to any one of claims 1 to 6.

Images