Personalized federal learning optimization method based on boundary reuse clustering

Abstract

The invention discloses a personalized federal learning optimization method based on boundary reuse clustering. The personalized federal learning optimization method comprises the following steps: S1, a central server initializes a global model and distributes the global model to each client; s2, updating the training model by the client; s3, the central server performs clustering by using a DBSCAN clustering algorithm, thereby determining the value range of the core data point; s4, dividing the boundary reuse data points into a plurality of different class clusters according to the overlapping degree between the boundary data and the core data points; s5, defining the client with the highest computing resource in each class cluster as a clustering center and collecting the clustering center; s6, the central server sets a weight according to the aggregated data size, so that a new global model obtained by aggregation of the weighted average method is issued to each client again; and S7, after the client receives the personalized hybrid model, training and updating are carried out again, and local training is carried out by using the personalized hybrid model so as to realize personalization.

Description

Technical Field

[0001] The present invention relates to the technical field of personalized federated learning and clustering algorithms, and in particular to a personalized federated learning optimization method based on boundary reuse clustering. Background Art

[0002] With the development of Internet technology, individuals and enterprises have generated a large amount of data. In many application scenarios, these data are scattered in different devices or institutions and are difficult to centralize. However, in traditional machine learning methods, in order to train an effective model, it is usually necessary to upload the data to a centralized server for processing. This practice often leads to the risk of user privacy leakage, and due to various restrictions on data transmission, difficulties such as "data islands" have gradually formed.

[0003] In order to meet this challenge, the federated learning method came into being; it allows multiple data sources to train models in parallel locally without sharing the original data information with the central server. This method has been widely studied for scenarios that protect user privacy and data security, such as its widespread use in mobile devices, banks, or smart healthcare; the core of federated learning is to keep the data on the client's terminal device and only send the model update parameters to the central server. Its biggest advantage is that the client does not need to directly share the original data, which effectively protects data privacy. This method not only solves data security and privacy issues, but also promotes multi-center collaboration, effectively improving the generalization ability and accuracy of the model.

[0004] Although federated learning has obvious advantages in protecting data privacy and promoting multi-party collaboration, a major challenge facing federated learning is the uneven distribution of data between clients across centers, which leads to uneven quality and accuracy of models trained on each device, thus affecting the accuracy and generalization ability of the global model. To solve this difficulty of data distribution differences, personalized federated learning (PFL) has gradually attracted the attention of the academic community. Personalized federated learning is an extended form developed on the basis of traditional federated learning. In reality, the data distribution of different clients may differ significantly, and this difference is called "heterogeneity". Personalized federated learning not only focuses on the global performance of the model, but also pays special attention to the personalized performance of the model on each client to adapt to the characteristics of its local data. By improving the model aggregation method, personalized federated learning enables the model to better adapt to the data distribution of different clients. Each client can adjust the global model according to the characteristics of its local data to generate a personalized model to adapt to the data characteristics of the local client.

[0005] The advantages of personalized federated learning are: Personalized federated learning can improve the robustness of the global model of the central server: Personalized federated learning reduces the impact of malicious clients on the global model by generating personalized models locally on the client. The central server can more easily identify and exclude abnormal client updates, thereby improving the robustness of the system; Personalized federated learning can improve the generalization of the global model of the central server: The central server can aggregate a more generalized global model by utilizing the data diversity of multiple clients. This is because the model of each client is optimized based on its local data. After receiving the models trained and updated by different clients, the central server aggregates the diverse model parameters to generate a more generalized global model; Personalized federated learning can better adapt to the data distribution of the client: After receiving the global model, each client can adjust it according to the local data to generate a personalized model. This local adjustment enables the model to better adapt to the data distribution of each client.

[0006] In large-scale personalized federated learning, frequent communication between all clients and the central server will lead to high communication overhead. At the same time, each client needs to perform complex calculations, which increases the burden on computing resources. In order to reduce computing overhead and communication costs, clustering algorithms have received a lot of attention. By clustering clients with similar data distributions together, clients in each cluster can share more information. After clustering, clients in each cluster can communicate locally more frequently, reducing the number of communications with the central server, thereby reducing communication overhead. However, existing clustering technologies divide the data points of the client into a cluster and do not fully consider whether the data points can be reused by the cluster to achieve the purpose of maximizing knowledge sharing. To address the shortcomings of the above-mentioned technologies, the present invention proposes a personalized federated learning optimization method based on boundary reuse clustering based on clustering algorithms. Summary of the invention

[0007] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a personalized federated learning optimization method based on boundary reuse clustering, which is used to solve the technical problem that the prior art does not fully consider whether data points can be reused by clusters to achieve the purpose of maximizing knowledge sharing.

[0008] To achieve the above object, the present invention provides a personalized federated learning optimization method based on boundary reuse clustering, comprising the following steps:

[0009] S1. The central server first initializes the weights, structure, and model parameters of the global model and distributes the initialized global model to each client through a secure communication channel;

[0010] S2. After receiving the initialized global model from the central server, the client updates the training model using the stochastic gradient descent (SGD) method according to the data distribution of the client's local data set.

[0011] S3, the central server uses the DBSCAN clustering algorithm to perform clustering, thereby determining the value range of the core data points;

[0012] S4, record the core data points of each client, construct a data distribution similarity function to determine the boundary data points, and divide the boundary reuse data points into several different clusters according to the degree of overlap between the boundary data and the core data points;

[0013] S5. The central server uses the global model to train all clients corresponding to the boundary reuse data points in the same cluster according to the category of the cluster, defines the client with the highest computing resources in each cluster as the cluster center and collects them;

[0014] S6. The clients represented by the cluster center and the updated training models of the clients are aggregated and uploaded to the central server. The central server sets weights according to the amount of aggregated data, so that the new global model obtained by weighted average aggregation is sent to each client again;

[0015] S7. The client sets a weight evaluation function, and the boundary reuse data points are introduced into the personalized hybrid model according to the weight evaluation function. After the client receives the personalized hybrid model, it is trained and updated again, and the personalized hybrid model is used for local training to achieve personalization.

[0016] Further: The specific method of S4 is as follows:

[0017] The data distribution of the core data points of each client is recorded as vector C = (C1, C2, ... C i , ...C n ), C n represents the core data point of cluster n, n is the total number of core data points of the client, (i, j)∈n; then and Another key data point and in Represents the core data point C i The distribution probability in the first data label; Represents the core data point C j The distribution probability in the first data label;

[0018] The data distribution of each client's boundary data point is recorded as vector B = (B1, B2, ... B i , ... B n), then and Another boundary data point and in Represents the boundary data point B i The distribution probability in the first data label; Represents the boundary data point B j The distribution probability in the first data label;

[0019] C i With C j are the core data points of different clusters determined by the DBSCAN clustering algorithm; here C i represents the core data point of cluster i, C j represents the core data point of cluster j, C i With C j Represents different core data points of different clusters. i and j are used to represent the indexes of different core data points, and i≠j;

[0020] Set the overlap threshold And calculate whether the distance between the boundary data point and the core data point in each client is less than the critical value of the overlap degree. If it meets the condition, the boundary data point is called a boundary reused data point and is divided into different clusters;

[0021] For each boundary data point B i And each core data point C i , the distance between them is calculated according to the probability vector of data label distribution, and the Euclidean distance is often used to calculate the similarity between two data points:

[0022]

[0023] when This indicates that the boundary data point B i At two different core data points C i and C j There is a certain degree of overlap between the distances, that is, B i exist Close to two different core data points C within the range of distance i and C j , then the boundary data point B i The reused data point for the boundary is denoted by Z i , and Z i At the same time, it is divided into core data points C i The clusters and core data points C j In the cluster composed of i belong clusters, among which λ is the total number of clusters. This processing method allows the client to identify data points located in the fuzzy area at the boundary of different clusters and assign them to the most relevant clusters, thereby promoting data knowledge sharing among different clients. At this time, data knowledge sharing among different clients can be maximized.

[0024] Further: The specific method of S5 is as follows:

[0025] The central server records the same type of cluster as u t , by each cluster u t The clients in the training jointly train the model and continuously iterate and update it until all the preset training rounds are completed;

[0026] The central server collects the computing power information of each client, and selects r cluster centers with high computing resources based on the client data information and the computing power score mapping table, denoted as r1, r2, r3, ...r n , i∈n, where the cluster center r i Clusters of the same type are aggregated to train updated models, and the model parameters aggregated by different clusters are uploaded to the central server as representatives.

[0027] Further: The specific method of S6 is as follows:

[0028] The central server receives data from cluster centers r1, r2, r3, ... r n Receive updated client local model The central server sets weights based on the amount of data and uses the weighted average method to aggregate and obtain a new global model. And send it to clients c1, c2, c3, ...c n , where the new global model The calculation formula is as follows:

[0029]

[0030] N is the total number of cluster centers, n r is the cluster center r i The amount of data, It is the model after cluster center aggregation update.

[0031] Further: The specific method of S7 is as follows:

[0032] Set a weight evaluation function W i (C) Method to evaluate and optimize the local model Introducing a personalized hybrid model The model is based on the client's weight evaluation function W i(C) The weight evaluation function expression is as follows:

[0033]

[0034] in Indicates the proportion of data point i to the total data volume, d i (C) represents the distance between data point i and the center of the cluster C to which it belongs, f(d i (C) The weight factor is used to reflect the impact of the distance between the data point and the cluster center on the contribution of the generated personalized model, that is, the weight evaluation between the amount of data and the distance. In order to ensure that the sum of the weights of all data points is 1, it is also necessary to add W i (C) Perform normalization; specifically, for each data point i belonging to clusters, and calculate their pairs The sum of the weights of the clusters is divided by the sum of the weights of the data point i to the total cluster λ;

[0035] For the boundary reuse data point Z i , which is the core data point C at the center of the cluster i The distance formula is as follows:

[0036]

[0037] The weight factor expression is:

[0038]

[0039] As the distance increases, f(d i (C)) will gradually decrease, and parameter a controls the decay rate;

[0040] The personalized hybrid model expression is:

[0041]

[0042] The model parameters of different clusters are combined by weighted averaging to generate a personalized model, where is the group of all clusters divided by the boundary reuse points Model parameters of

[0043] Boundary reuse data point z i Adjust the value of a according to the needs to evaluate its own contribution to optimize the performance of the local model. The client generates a personalized hybrid model based on the data point contribution evaluation. Achieve the goal of personalized federated learning;

[0044] After the client receives the personalized hybrid model, it is trained and updated again. At this time, the client uses the personalized hybrid model Perform local training to achieve personalization. The training results are as follows:

[0045]

[0046] The beneficial effect of the present invention is that while reducing the communication overhead, it also fully considers the property that data points can be reused by clusters to achieve the purpose of maximizing knowledge sharing. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention and the technical solutions in the prior art, the drawings required for use in the embodiments and the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some implementation methods recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0050] Obviously, the drawings described below are only some examples or embodiments of the present invention. For ordinary technicians in this field, the present invention can be applied to other similar scenarios based on these drawings without any creative work.

[0051] In addition, it can be understood that although the efforts made in this development process may be complex and lengthy, for ordinary technicians in the field related to the contents disclosed in the present invention, some design, manufacturing or production changes based on the technical contents disclosed in the present invention are just conventional technical means and should not be understood as the contents disclosed in the present invention are insufficient.

[0052] If not otherwise specified, all embodiments and optional embodiments of the present invention can be combined with each other to form a new technical solution.

[0053] See also Figure 1 , the method flow diagram of the present invention comprises the following steps:

[0054] S1. The central server first initializes the weights, structure, and model parameters of the global model and distributes the initialized global model to each client through a secure communication channel;

[0055] S2. After receiving the initialized global model from the central server, the client updates the training model using the stochastic gradient descent (SGD) method according to the data distribution of the client's local data set.

[0056] S3, the central server uses the DBSCAN clustering algorithm to perform clustering, thereby determining the value range of the core data points;

[0057] S4, record the core data points of each client, construct a data distribution similarity function to determine the boundary data points, and divide the boundary reuse data points into several different clusters according to the degree of overlap between the boundary data and the core data points;

[0058] S5. The central server uses the global model to train all clients corresponding to the boundary reuse data points in the same cluster according to the category of the cluster, defines the client with the highest computing resources in each cluster as the cluster center and collects them;

[0059] S6. The clients represented by the cluster center and the updated training models of the clients are aggregated and uploaded to the central server. The central server sets weights according to the amount of aggregated data, so that the new global model obtained by weighted average aggregation is sent to each client again;

[0060] S7. The client sets a weight evaluation function, and the boundary reuse data points are introduced into the personalized hybrid model according to the weight evaluation function. After the client receives the personalized hybrid model, it is trained and updated again, and the personalized hybrid model is used for local training to achieve personalization.

[0061] Specifically:

[0062] The weights, structure and model parameters of the global model first initialized by the central server are recorded as The initialized global model is distributed to each client through a secure communication channel to ensure the privacy and security of the initialized model. After receiving the initialized global model from the central server, the client records the initial model of each client data as Divide the data D into s different batches of data size S. For each batch of data D s , the client uses the stochastic gradient descent method to update the training model, and stops training when the client model parameters reach the preset accuracy;

[0063] On the nth client c n In , after each federated learning round t:

[0064] For data D s The local model updates are:

[0065] For data D s The global model updates are:

[0066] in, For client c n The local model, is the global model of the central server, η is the learning rate, is the loss of the model on the training samples.

[0067] The central server sets the clustering parameters ε and M P The value determines the neighborhood of clustering and the number of minimum object points. When the distance of a data point is less than the neighborhood radius and the number of data points in the neighborhood of the data point is greater than or equal to the number of minimum object points, a new cluster is created for the data point, as shown below:

[0068] dist(C x )≤ε,num(C x )≥MP;

[0069] C x is the data point, C x The data point that satisfies the above conditions is the core data point. Find the data point C x Other data points in the neighborhood are added to the cluster until all data points have been visited; if a data point does not meet the core data point requirements but is in the middle of the neighborhood radius of the core data point, the data point is called a boundary data point; the remaining data points that neither meet the core data point requirements nor meet the neighborhood radius of the core data point are noise points, thus classifying boundary data points and noise points according to the distribution of data points in the neighborhood.

[0070] Then design the boundary function and calculate the data distribution vectors of the core data points and boundary data points of each client, determine the boundary reuse data points and divide them into corresponding clusters. The method is: the data distribution of the core data points of each client is recorded as vector C = (C1, C2, ... C i , ...C n ), C n represents the core data point of cluster n, n is the total number of core data points of the client, (i, j)∈n; then and Another key data point and in Represents the core data point C i The distribution probability in the first data label; Represents the core data point Cj The distribution probability in the first data label;

[0071] The data distribution of each client's boundary data point is recorded as vector B = (B1, B2, ... B i , ... B n ), then and Another boundary data point and in Represents the boundary data point B i The distribution probability in the first data label; Represents the boundary data point B j The distribution probability in the first data label;

[0072] C i With C j are the core data points of different clusters determined by the DBSCAN clustering algorithm; here C i represents the core data point of cluster i, C j represents the core data point of cluster j, C i With C j Represents different core data points of different clusters. i and j are used to represent the indexes of different core data points, and i≠j;

[0073] Set the overlap threshold And calculate whether the distance between the boundary data point and the core data point in each client is less than the critical value of the overlap degree. If it meets the condition, the boundary data point is called a boundary reused data point and is divided into different clusters;

[0074] For each boundary data point B i And each core data point C i , the distance between them is calculated according to the probability vector of data label distribution, and the Euclidean distance is often used to calculate the similarity between two data points:

[0075]

[0076] when This indicates that the boundary data point B i At two different core data points C i and C j There is a certain degree of overlap between the distances, that is, B i exist Close to two different core data points C within the range of distance i and C j , then the boundary data point B i The reused data point for the boundary is denoted by Zi , and Z i At the same time, it is divided into core data points C i The clusters and core data points C j In the cluster composed of i belong clusters, among which λ is the total number of clusters. This processing method allows the client to identify data points located in the fuzzy area at the boundary of different clusters and assign them to the most relevant clusters, thereby promoting data knowledge sharing among different clients. At this time, data knowledge sharing among different clients can be maximized.

[0077] The central server marks the clusters of the same type, and after taking values ​​for the clients in each cluster, the values ​​are used to train the model together; the central server marks the same cluster as u t , by each cluster u t The clients in the training jointly train the model and continuously iterate and update it until all the preset training rounds are completed;

[0078] The central server collects the computing power information of each client, and selects r cluster centers with high computing resources based on the client data information and the computing power score mapping table, denoted as r1, r2, r3, ...r n , i∈n, where the cluster center r i Clusters of the same type are aggregated to train updated models, and the model parameters aggregated by different clusters are uploaded to the central server as representatives.

[0079] The central server receives data from cluster centers r1, r2, r3, ... r n Receive updated client local model The central server sets weights based on the amount of data and uses the weighted average method to aggregate and obtain a new global model. And send it to clients c1, c2, c3, ...c n , where the new global model The calculation formula is as follows:

[0080]

[0081] N is the total number of cluster centers, n r is the cluster center r i The amount of data, It is the model after cluster center aggregation update.

[0082] Set a weight evaluation function W i (C) Method to evaluate and optimize the local model Introducing a personalized hybrid model The model is based on the client's weight evaluation function W i (C) The weight evaluation function expression is as follows:

[0083]

[0084] in Indicates the proportion of data point i to the total data volume, d i (C) represents the distance between data point i and the center of the cluster C to which it belongs, f(d i (C) The weight factor is used to reflect the impact of the distance between the data point and the cluster center on the contribution of the generated personalized model, that is, the weight evaluation between the amount of data and the distance. In order to ensure that the sum of the weights of all data points is 1, it is also necessary to add W i (C) Perform normalization; specifically, for each data point i belonging to clusters, and calculate their pairs The sum of the weights of the clusters is divided by the sum of the weights of the data point i to the total cluster λ;

[0085] For the boundary reuse data point Z i , which is the core data point C at the center of the cluster i The distance formula is as follows:

[0086]

[0087] The weight factor expression is:

[0088]

[0089] As the distance increases, f(d i (C)) will gradually decrease, and parameter a controls the decay rate;

[0090] The personalized hybrid model expression is:

[0091]

[0092] The model parameters of different clusters are combined by weighted averaging to generate a personalized model, where It is a cluster Model parameters of

[0093] Boundary reuse data point z i Adjust the value of a according to the needs to evaluate its own contribution to optimize the performance of the local model. The client generates a personalized hybrid model based on the data point contribution evaluation. Achieve the goal of personalized federated learning;

[0094] After the client receives the personalized hybrid model, it is trained and updated again. At this time, the client uses the personalized hybrid model Perform local training to achieve personalization. The training results are as follows:

[0095]

[0096] Aiming at the problem of low knowledge sharing rate of client data in federated learning, the present invention proposes a personalized federated learning optimization method based on boundary reuse clustering;

[0097] This method first performs effective clustering based on the client data distribution information, so that clients in the same cluster have similar data requirements, thereby reducing the data resource computing cost of a single client;

[0098] Design boundary functions to determine boundary reuse points, and then assign different boundary reuse points to different clusters. This design greatly improves the utilization of data resources and can effectively improve the efficiency of knowledge sharing between different clusters. After the global model is aggregated, it is sent to each client. The boundary reuse data points first evaluate the contribution of the data points. The client generates a personalized model based on the contribution evaluation of the data points to optimize the performance of the local model and meet the needs of personalized federated learning for different clients.

[0099] After the personalized model is finally evaluated, the client receives the personalized hybrid model and uses the local data set to train and update the model until the preset accuracy is reached.

[0100] The above is a very detailed description of one or more embodiments of the present invention, but the content is only a specific example of the present invention and cannot be considered to limit the scope of application of the present invention. Any other methods and changes proposed based on the content of the present invention should fall within the scope of patent protection of the present invention.

Claims

1. A personalized federated learning optimization method based on boundary reuse clustering, characterized in that: The steps include: S1. The central server first initializes the weights, structure, and model parameters of the global model and distributes the initialized global model to each client through a secure communication channel; S2. After receiving the initialized global model from the central server, the client updates the training model using the stochastic gradient descent (SGD) method according to the data distribution of the client's local data set. S3, the central server uses the DBSCAN clustering algorithm to perform clustering, thereby determining the value range of the core data points; S4, record the core data points of each client, construct a data distribution similarity function to determine the boundary data points, and divide the boundary reuse data points into several different clusters according to the degree of overlap between the boundary data and the core data points; S5. The central server uses the global model to train all clients corresponding to the boundary reuse data points in the same cluster according to the category of the cluster, defines the client with the highest computing resources in each cluster as the cluster center and collects them; S6. The clients represented by the cluster center and the updated training models of the clients are aggregated and uploaded to the central server. The central server sets weights according to the amount of aggregated data, so that the new global model obtained by weighted average aggregation is sent to each client again; S7. The client sets a weight evaluation function, and the boundary reuse data points are introduced into the personalized hybrid model according to the weight evaluation function. After the client receives the personalized hybrid model, it is trained and updated again, and the personalized hybrid model is used for local training to achieve personalization.

2. According to claim 1, a personalized federated learning optimization method based on boundary reuse clustering is characterized in that: The specific method of S2 is as follows: The initial model of each client data is Divide the data D into s different batches of data size S. For each batch of data D s , the client uses the stochastic gradient descent method to update the training model, and stops training when the client model parameters reach the preset accuracy; On the nth client c n In , after each federated learning round t: For data D s The local model updates are: For data D s The global model updates are: in, For client c n The local model, is the global model of the central server, η is the learning rate, is the loss of the model on the training samples.

3. According to claim 1, a personalized federated learning optimization method based on boundary reuse clustering is characterized in that: The specific method of S3 is as follows: The central server sets the clustering parameters ε and M P The value determines the neighborhood of clustering and the number of minimum object points. When the distance of a data point is less than the neighborhood radius and the number of data points in the neighborhood of the data point is greater than or equal to the number of minimum object points, a new cluster is created for the data point, as shown below: dist(C x )≤ε,num(C x )≤M P ; C x is the data point, C x The data point that satisfies the above conditions is the core data point. Find the data point C x Other data points in the neighborhood are added to the cluster until all data points have been visited; if a data point does not meet the core data point requirements but is in the middle of the neighborhood radius of the core data point, the data point is called a boundary data point; the remaining data points that neither meet the core data point requirements nor meet the neighborhood radius of the core data point are noise points.

4. According to claim 1, a personalized federated learning optimization method based on boundary reuse clustering is characterized in that: The specific method of S4 is as follows: The data distribution of the core data points of each client is recorded as vector C = (C1, C2, ... C i , ...C n ), C n represents the core data point of cluster n, n is the total number of core data points of the client, (i, j)∈n; then and Another key data point and in Represents the core data point C i The distribution probability in the first data label; Represents the core data point C j The distribution probability in the first data label; The data distribution of each client's boundary data points is recorded as vector B=(B1,B2,...B i ,...B n ), then and Another boundary data point and in Represents the boundary data point B i The distribution probability in the first data label; Represents the boundary data point B j The distribution probability in the first data label; C i With C j are the core data points of different clusters determined by the DBSCAN clustering algorithm; here C i represents the core data point of cluster i, C j represents the core data point of cluster j, C i With C j Represents different core data points of different clusters. i and j are used to represent the indexes of different core data points, and i≠j; Set the overlap threshold And calculate whether the distance between the boundary data point and the core data point in each client is less than the critical value of the overlap degree. If it meets the condition, the boundary data point is called a boundary reused data point and is divided into different clusters; For each boundary data point B i And each core data point C i , the distance between them is calculated according to the probability vector of data label distribution, and the Euclidean distance is often used to calculate the similarity between two data points: when This indicates that the boundary data point B i At two different core data points C i and C j There is a certain degree of overlap between the distances, that is, B i exist Close to two different core data points C within the range of distance i and C j , then the boundary data point B i The reused data point for the boundary is denoted by Z i , and Z i At the same time, it is divided into core data points C i The clusters and core data points C j In the cluster composed of i belong clusters, among which λ is the total number of clusters. This processing method allows the client to identify data points located in the fuzzy area at the boundary of different clusters and assign them to the most relevant clusters, thereby promoting data knowledge sharing among different clients. At this time, data knowledge sharing among different clients can be maximized.

5. The personalized federated learning optimization method based on boundary reuse clustering according to claim 1, characterized in that: The specific method of S5 is as follows: The central server records the same type of cluster as u t , by each cluster u t The clients in the training jointly train the model and continuously iterate and update it until all the preset training rounds are completed; The central server collects the computing power information of each client, and selects r cluster centers with high computing resources based on the client data information and the computing power score mapping table, denoted as r1, r2, r3, ...r n , i∈n, where the cluster center r i Clusters of the same type are aggregated to train updated models, and the model parameters aggregated by different clusters are uploaded to the central server as representatives.

6. The personalized federated learning optimization method based on boundary reuse clustering according to claim 1, characterized in that: The specific method of S6 is as follows: The central server receives data from cluster centers r1, r2, r3, ... r n Receive updated client local model The central server sets weights based on the amount of data and uses the weighted average method to aggregate and obtain a new global model. And send it to clients c1, c2, c3, ...c n , where the new global model The calculation formula is as follows: N is the total number of cluster centers, n r is the cluster center r i The amount of data, It is the model after cluster center aggregation update.

7. The personalized federated learning optimization method based on boundary reuse clustering according to claim 1, characterized in that: The specific method of S7 is as follows: Set a weight evaluation function W i (C) Method to evaluate and optimize the local model Introducing a personalized hybrid model This model is based on the client's weight evaluation function W i (C) The weight evaluation function expression is as follows: in Indicates the proportion of data point i to the total data volume, d i (C) represents the distance between data point i and the center of the cluster C to which it belongs, f(d i (C) The weight factor is used to reflect the impact of the distance between the data point and the cluster center on the contribution of the generated personalized model, that is, the weight evaluation between the amount of data and the distance. In order to ensure that the sum of the weights of all data points is 1, it is also necessary to add W i (C) Perform normalization; specifically, for each data point i belonging to clusters, and calculate their pairs The sum of the weights of the clusters is divided by the sum of the weights of the data point i to the total cluster λ; For the boundary reuse data point Z i , which is the core data point C at the center of the cluster i The distance formula is as follows: The weight factor expression is: As the distance increases, f(d i )C)) will gradually decrease, and parameter a controls the decay rate; The personalized hybrid model expression is: The model parameters of different clusters are combined by weighted averaging to generate a personalized model, where is the group of all clusters divided by the boundary reuse points Model parameters of Boundary reuse data point Z i Adjust the value of a according to the needs to evaluate its own contribution to optimize the performance of the local model. The client generates a personalized hybrid model based on the data point contribution evaluation. Achieve the goal of personalized federated learning; After the client receives the personalized hybrid model, it is trained and updated again. At this time, the client uses the personalized hybrid model Perform local training to achieve personalization. The training results are as follows:

Images