Fair graph federated learning method based on contribution evaluation

Abstract

The invention discloses a fair map federated learning method based on contribution evaluation, which evaluates the contribution of an agent and adopts a fair distribution strategy, thereby improving the fairness and efficiency of federated learning. The method specifically comprises the following steps: calculating a contribution value of each agent to federated learning according to a local gradient uploaded by each agent and graph diversity of local data; integrating the local gradients uploaded by the agents by using a weighted average method to obtain a global gradient for training a centralized graph model; according to a graph structure mode contained in local data of each agent, calculating a graph prototype between each agent and the server, and taking the graph prototype as a reference coordinate for communication; performing sparse processing on the global gradient according to the contribution value of each agent to obtain model parameter configuration received by each agent; and calculating a feedback score received by each agent according to the contribution value of each agent in the historical round, and carrying out suppression processing or correction on the agents with harmful or delayed contributions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of graph federated learning, and particularly relates to a fair graph federated learning method based on contribution evaluation. BACKGROUND

[0002] In the current information age, large-scale graph data is widely used in various fields, including social network analysis, recommendation systems, bioinformatics, etc. These graph data are often stored by multiple data holders (such as different organizations, institutions, individuals) and contain rich information and potential knowledge. However, due to privacy and security restrictions, data holders are often reluctant to centrally aggregate their sensitive data to a single central server for modeling and analysis. Current federated learning methods can achieve model training in distributed data, but for structured data such as graph data, traditional federated learning methods have some challenges. First, the structural nature of graph data requires accurate modeling of node and edge connection relationships, and traditional federated learning methods lack explicit modeling capabilities for graph structures. Second, graph data often has high dimensionality and sparsity, which makes it difficult for traditional federated learning methods to effectively utilize the information of graph data. Researchers have developed methods such as FedSage, GCFL, and FedStar to conduct graph federated learning. Among them, FedSage replaces the encoder in the FedAvg method with GraphSage to better capture graph structure information; GCFL uses hierarchical clustering methods to aggregate agents according to gradients; FedStar only federates the structural information of graph data, while the feature information is only learned locally. However, none of these methods takes into account the fact that federated learning participants are selfish and may not be willing to contribute all their data for free. There must be a fair model performance feedback to ensure that the computational resource investment matches the model benefit, otherwise they may quit the federated learning.

[0003] The fairness of federated learning has only been studied on image data, but on graphs, which are non-Euclidean data, there may be a negative contribution effect, that is, the model trained by the data of some participants may have a negative effect on the global model; even if the graphs come from the same dataset, this situation may occur. Therefore, the fairness of graph federated learning needs to be studied. SUMMARY

[0004] In order to make up for the shortcomings of the prior art, the application aims to provide a fair graph federated learning method based on contribution evaluation, which realizes effective sharing and integration of multi-party data while protecting data privacy by jointly modeling and learning graph data in a distributed environment, thereby promoting the collaborative training of machine learning models and the improvement of generalization ability.

[0005] The fair graph federated learning method based on contribution evaluation includes the following steps: Step 1: Calculate the contribution of each agent to federated learning based on the local gradients uploaded by each agent and the graph diversity of the local data. Step 2: Based on the contribution value of each agent, use a weighted average method to integrate the local gradients uploaded by each agent to obtain the global gradient, which is used to train the centralized graph model. Step 3: Based on the graph structure pattern contained in the local data of each agent, calculate the graph prototype between each agent and the server, and use it as a reference coordinate for communication; Step 4: Based on the contribution value of each agent, perform sparsification on the global gradient to obtain the model parameter configuration received by each agent; Step 5: Calculate the feedback score received by each agent based on their contribution value in the historical rounds, and suppress or correct agents with negative migration risk or lagging updates.

[0006] Furthermore, in step 1, the contribution value is calculated using the Shapley value estimation method, employing locally uploaded gradients. With global gradient The marginal contribution is estimated using the cosine similarity, and the specific calculation formula is as follows: ,in, It is the number of participants. For a client-side group, For all possible combinations, Let's define the similarity function as follows: To measure the consistency between local and global gradients, the contribution value is evaluated using an exponential moving average method, combined with historical contribution values. and current contribution value Through parameters Balancing the weights of historical and current information can be seen as a dynamic weighting method that ensures the accuracy and real-time nature of agent contribution values.

[0007] Graph diversity is measured by the number of motifs, which is determined by counting the number of different motifs held by each participant. The diversity index is obtained by normalizing the diversity index relative to the total number K of base graphs held by all participants. ,in, It is the first The number of base graphs held by each participant, K is the sum of the number of base graphs held by all participants; multiplying graph diversity by marginal contribution yields the evaluation value of each participant for the entire federated framework at each round. .

[0008] Further, in step 2, the evaluation value of the current participant is calculated by using the weighted average of the historical round and the current round, and the calculation formula is: , wherein, represents the evaluation value of the first participant in the first round, is the weighting parameter of the forward round and the current round evaluation value.

[0009] Further, in step 3, the participant evaluation value is normalized, so that the sum of the evaluation values of each participant is 1, and the calculation formula is: , and let as the initial evaluation value of each participant.

[0010] Further, in step 4, the sparse processing adopts mask operation, and each participant is distributed with sparse model parameter configuration, and the specific calculation formula is , wherein, operation returns the largest elements in the original gradient D is the number of elements in the gradient, is a pre-set hyperparameter.

[0011] Further, in step 5, according to the contribution value of each agent, the feedback score is proportionally distributed, accounting for the proportion of the system reference total amount; for the participants who are harmful to the overall federation, that is, the participants with a contribution value , give negative feedback score as inhibition processing, represented as , deduct the score from the participant; for the participants who have positive effect on the overall federation, that is, the participants with , the basic score is proportionally distributed based on the contribution value .

[0012] Further, in step 5, the calculation formula of the lagging update correction amount is , which calculates the difference between the current round contribution value and the historical average contribution value, and gives an additional correction gain when the current contribution is significantly higher than the historical average.

[0013] Compared with the prior art, the present application has the following advantages: (1) The present application maximizes fairness without sacrificing model performance. The fairness here includes the fairness of model distribution and the fairness of participant feedback score, which means that each participant can enjoy the results of federated learning fairly, and the overall model performance can be maintained.

[0014] (2) The present application represents an important progress in realizing fairness in graph federated learning. By accurately evaluating the contribution of agents and adopting a fair allocation strategy, the fairness and efficiency of federated learning are improved.

[0015] (3) The core problem solved by the present application is not "how to allocate benefits", but two technical difficulties inherent in graph federated learning: ① Negative transfer problem: In distributed training, the difference in data distribution (Non-IID) of some nodes will lead to a decline in the performance of the global model; the present application identifies "negative contribution" nodes and performs "suppression processing" (i.e. gradient truncation or weight zeroing), eliminating noise interference from the perspective of signal processing to improve the convergence accuracy of the model.

[0016] ② Communication bandwidth limited problem: The communication cost is high in federated learning; the present application performs "sparse processing" (Mask operation) on the gradient according to the "contribution evaluation" result, dynamically adjusting the amount of parameters issued; this is a data compression and communication resource scheduling technology that can save network bandwidth. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of the present application; Figure 2 is the allocation fairness experimental result of the experimental verification of the present application. DETAILED DESCRIPTION

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

[0019] The present application provides a fair graph federated learning method based on contribution evaluation, which can protect the privacy of local graph data of multiple agents (agents) while collaboratively training a centralized graph model. First, the global prototype is initialized; then, the federated learning operation is performed, and the communication is performed several times (i.e. ① aggregate the local gradient, ② update the evaluation value of each participant, ③ update locally, and ④ update the local prototype and the global prototype). As shown in Figure 1 The specific steps of the learning method are as follows: (1) Step 1, according to the local gradient uploaded by each agent and the graph diversity of local data, calculate the contribution value of each agent to federated learning.

[0020] The Shapey value is used to calculate the marginal contribution of each participant to the entire aggregation process when performing federated learning aggregation as follows: , Where N is the number of participants. For a client-side group, For all possible combinations, Let be the similarity function, defined here as This refers to the cosine similarity between the gradient uploaded by group S and the global gradient. This method allows for accurate evaluation of gradient consistency among each participant in federated learning, providing evaluable information for subsequent federated frameworks. This facilitates a more equitable allocation of federated learning tasks and computational resources.

[0021] Here we can see that calculating this Shapley value requires high complexity, that is... Therefore, this application employs an estimation method, using locally uploaded gradients. With global gradient The marginal contribution is estimated using the cosine similarity. Furthermore, it can be proven that the gap between these two factors can be controlled.

[0022] It can be further observed that graph diversity can also influence the degree to which participants contribute to the overall federated framework. Therefore, this invention uses the number of base graphs to measure graph diversity: , in, Let K be the number of base graphs held by the i-th participant, and K be the sum of the number of base graphs held by all participants. This refers to the graph diversity as defined in this invention. Multiplying the graph diversity by the marginal contribution yields the evaluation value of each participant for the entire federated framework in each round, i.e.: .

[0023] (ii) Step 2: Based on the contribution value of each agent, use the weighted average method to integrate the local gradients uploaded by each agent to obtain the global gradient, which is used to train the centralized graph model.

[0024] The current participant's evaluation score is calculated using a weighted average of historical rounds and the current round. The formula is as follows: ,in, This represents the evaluation value of the i-th participant in the t-th round. This is a parameter that balances the evaluation value of the previous round and the current round.

[0025] Contribution values ​​are used as weights for gradient aggregation, unlike traditional federated learning methods which rely solely on the size of the local training set. Gradient aggregation employs a weighted average method, using participant evaluations as weights, with the specific calculation formula as follows: ,in, For the first The local gradient uploaded by each participant in round t. As the weight of the participant's evaluation value, This represents the aggregated global gradient; this weight allocation method more accurately reflects the agent's actual contribution to federated learning.

[0026] Due to the unique nature of graph federated learning, some agents may have negative contribution values, indicating a potential negative transfer to the overall federated task. Therefore, to avoid these negative impacts on the entire federated learning task, it is necessary to exclude agents with negative contribution values ​​from the federated framework. The proposed method uses the ReLU activation function to process the participants' evaluation values, i.e. Participants with negative contributions are excluded from gradient aggregation, ensuring that only participants with positive contributions participate in the construction of the global gradient; the aggregation weights are normalized based on the evaluation values ​​after ReLU processing to ensure the numerical stability of the global gradient.

[0027] (III) Step 3: Based on the graph structure pattern contained in the local data of each agent, i.e. the base graph (motif), calculate the graph prototype between each agent and the server, and use it as a reference coordinate for communication.

[0028] The participants' evaluation scores were normalized so that the sum of each participant's evaluation score was 1. The calculation formula is as follows: and order This serves as the initial evaluation value for each participant.

[0029] This invention also aggregates the graph prototypes to further improve the quality of the local model. In this invention, for a base graph k on a participant i, the local prototype is defined as: , in, This indicates that a training instance of base graph k is contained on participant i. This represents the model parameters for participant i in round t. Then, the central client aggregates the local prototypes of base graph k to obtain the global prototype, i.e.: , in, Represents the set of participants containing the base graph k. This represents the number of all graph instances containing the base graph k. In this invention, participants upload their local graph prototypes to the server, which aggregates and distributes them before updating their local models. This achieves effective knowledge sharing and helps to more accurately adjust the local models to better meet the needs of the global model. For updating the local model, this invention defines the loss function as follows: , in, This refers to the output on the local model. Supervision loss between label Y, It's a balance between the supervised loss and the regularization term; the distance metric d used here is... distance.

[0030] (iv) Step 4: Based on the contribution value of each agent, perform sparsification on the global gradient to obtain the model parameter configuration received by each agent.

[0031] The gradient is sparsified based on the agent's contribution value in the current round. This means that agents with high contribution values ​​will receive gradients more similar to the global gradient, which helps improve the starting point of their model learning. The model sparsification uses a mask operation to distribute sparse model parameters to each participant. The specific calculation formula is as follows: ,in, The operation returns the original gradient. The largest x elements in the gradient, where D is the number of elements in the gradient. These are pre-set hyperparameters; this strategy can be viewed as a dynamic, contribution-based parameter configuration mechanism.

[0032] The degree of sparsity is determined by the contribution values ​​of the participants. The decision is to use the tanh function for normalization to ensure that participants with high contribution values ​​receive more complete gradient information, while avoiding the distribution of inverse gradients to participants, so as not to impair the model performance of federated learning; this helps to balance the contributions and model performance among different agents in federated learning.

[0033] (v) Step 5: Calculate the feedback score received by each agent based on the contribution value of each agent in the historical rounds, and suppress or correct agents that make harmful or delayed contributions.

[0034] Based on each agent's contribution, their feedback scores are allocated proportionally to the total system reference score, ensuring that agents receive corresponding score weights according to their contributions and reflecting the fairness of distributed learning.

[0035] For participants who are detrimental to the overall federation (i.e., contribution value) (Participants): This invention assigns negative feedback scores to these participants, i.e., performs certain inhibitory treatments on them, denoted as... .

[0036] For participants who have a positive effect on the overall federation (i.e., contribution value) (Participants): The baseline ratings of these participants are based on their assessment values. The calculation is performed. Furthermore, considering the delayed feedback characteristic of graph federation learning, this invention also provides some correction gains for participants experiencing delayed feedback, expressed as hysteresis correction update amounts. : , The formula calculates the difference between the current round's contribution value and the historical average contribution value, and provides an additional correction gain when the current contribution is significantly higher than the historical average. The correction mechanism takes into account the delayed feedback characteristics that may exist in graph federated learning, that is, the contributions of some participants may only be reflected in subsequent rounds. The correction mechanism ensures that these participants receive fair rewards. The final feedback score of the participants is the base score plus the lag update correction amount, which helps to maintain the active participation of agents and maintain the efficiency and fairness of the federated learning task.

[0037] The technical effects of the present invention are verified through experiments below.

[0038] (1) Datasets. In this embodiment, the model performance was validated on three datasets (as shown in Table 1), which included various types of graphs (social networks, molecular networks) and their average edge sizes ranged from tens to hundreds.

[0039] Table 1: Information on nodes, edges, average degree, etc. of the experimental graphs in the three datasets.

[0040] This invention compares seven commonly used models with the model proposed in this application. Here, ours refers to the model proposed in this invention. Among the seven models used in the comparative experiment, FedAvg is a classic federated learning algorithm; FedProx is the first algorithm proposed for heterogeneous federated learning; FedSage is an algorithm that transforms the local encoder of FedAvg into GraphSage; GCFL and GCFL+ are algorithms that perform clustering based on locally uploaded gradients and are the most advanced graph federated learning algorithms;

[33] is a fair federated learning algorithm proposed on image data.

[0041] Table 2: Experimental results for the three original datasets

[0042] (2) Accuracy test This invention achieves the best fairness in model allocation with minimal performance sacrifice. Simultaneously, it also achieves the best overall model accuracy.

[0043] (3) Distributive fairness experiment This invention first perturbs the edges of the graph data for each participant to varying degrees. The lower the perturbation ratio, the higher the data quality of the graph data. The perturbation ratio is defined as [0,0.3] as "high", [0.3,0.7] as "medium", and [0.7,1] as "low".

[0044] This invention compares four common contribution assessment allocation schemes: (1) Union scoring allocation mechanism: considering the difference in scores before and after removing a participant as a marginal contribution measure; (2) Shapley scoring allocation mechanism: calculating the Shapley value as a marginal contribution measure; (3) Individual scoring allocation mechanism: calculating only the individual score of a participant; (4) Equal allocation mechanism: each participant has an equal marginal score. Figure 2 As shown, the present invention can give the highest feedback score to the participant with the highest data quality and the negative score to the participant with low data quality during the allocation process, indicating that the present invention can have good scoring fairness.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fairness graph federated learning method based on contribution evaluation, characterized in that, The method comprises the following steps: Step 1, calculating the contribution value of each agent to federated learning according to the local gradient uploaded by each agent and the graph diversity of local data; Step 2, based on the contribution value of each agent, using the weighted average method to integrate the local gradient uploaded by each agent to obtain the global gradient for training the centralized graph model; Step 3, calculating the graph prototype between each agent and the server according to the graph structure mode contained in the local data of each agent, and using the graph prototype as the reference coordinate for communication; Step 4, according to the contribution value of each agent, the global gradient is sparsified to obtain the model parameter allocation amount received by each agent; Step 5, according to the contribution value of each agent in the historical round, the feedback score received by each agent is calculated, and the agents with harmful or delayed contribution are inhibited or corrected.

2. The fair graph federated learning method based on contribution evaluation according to claim 1, characterized in that, In step 1, the calculation of the contribution value adopts the estimation method of Shapley value, using the locally uploaded gradient Cosine similarity with global gradient to estimate its marginal contribution, the specific calculation formula is: Wherein, is the number of participants, is a client team, is all possible combinations, is the similarity function, defined as ; The graph diversity adopts the number of base graphs to measure, the calculation formula is: Wherein, is the number of base graphs held by the th participant, is the sum of the number of base graphs held by all participants; Multiply the graph diversity by the marginal contribution to get the evaluation value of each participant in each round for the entire federal framework .

3. The fair graph federated learning method based on contribution evaluation according to claim 2, characterized in that, In step 2, the evaluation value of the current participant is calculated by using the weighted average of the historical round and the current round, and the calculation formula is: Wherein, represents the evaluation value of the i th participant in the j th round, is the weighting parameter of the forward round and the current round evaluation value.​​ 4. The fair graph federated learning method based on contribution evaluation according to claim 3, characterized in that, The evaluation value of each participant is normalized in step 3, so that the sum of the evaluation values of all participants is 1, and the calculation formula is: Let be the initial evaluation value of each participant.

5. The fair graph federated learning method based on contribution evaluation according to claim 3, characterized in that, In step 4, the sparsification process adopts a mask operation to distribute the sparsified model parameter configuration to each participant, and the specific calculation formula is wherein, The operation returns the original gradient The maximum elements in The number of elements in the gradient is a pre-set hyperparameter.

6. The fair graph federated learning method based on contribution evaluation according to claim 3, characterized in that, In step 5, the feedback score is proportionally distributed according to the contribution value of each agent, in proportion to the total amount of system reference; for participants who are harmful to the overall federation, i.e. participants with a contribution value , a negative feedback score is given as a suppression treatment, denoted as , and the score is deducted from the participant; for participants who have a positive effect on the overall federation, i.e. participants with a contribution value , the base score is proportionally distributed based on the contribution value .

7. The fair graph federated learning method based on contribution evaluation according to claim 3, characterized in that, The step 5, the lag update correction amount The calculation formula is The formula calculates the difference between the current round contribution value and the historical average contribution value, and gives an additional correction gain when the current contribution is significantly higher than the historical average.

Images