Personalized federal map learning method and system based on depth map similarity learning

Abstract

The invention belongs to the field of privacy calculation and data security, and particularly relates to a personalized federated map learning method and system based on depth map similarity learning, and the method comprises the steps: collecting local map data, and obtaining an unsupervised map encoder through transverse federated average learning; inputting local graph data into an unsupervised graph encoder, and calculating a dense vector; inputting the dense vector into a central server to calculate the similarity, and constructing a personalized model; performing federal iteration training on the participants according to the personalized model; the method is composed of two core parts, namely a personalized parameter aggregation mechanism based on depth map similarity learning and a participant selection mechanism based on uniform label distribution, an unsupervised map encoder is trained cooperatively, and low-dimensional dense vectors are generated based on local map data of participants. Therefore, the model actually obtained by the participant can adapt to the local data, so that the data actually participating in the training is distributed as uniformly as possible, the representation of the information is more comprehensive, and the universality and generalization of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the field of privacy computing and data security, and specifically relates to a personalized federated graph learning method and system based on deep graph similarity learning. Background Art

[0002] Graph neural networks can extract features and learn embedded representations from complex graph structure data in the real world, and have a wide range of applications in various fields. The training of graph neural network models requires big data as a basic condition. However, as society pays more and more attention to personal privacy and data security, various privacy-related laws and regulations have been introduced at home and abroad to strengthen the supervision of data privacy, especially for structured data such as graph data. For individuals, big data usually contains sensitive information such as personal identity, interests, and geographic location, which may bring security risks once leaked; for enterprises, data is one of their lifebloods, and business data usually involves commercial secrets. Once leaked, it may cause a devastating blow to business operations. Therefore, in reality, graph data is usually distributed in the hands of multiple data holders themselves. Due to privacy issues, it is impossible to directly centralize it for storage and use, thus forming many "data islands", making it difficult to circulate graph data and release its value.

[0003] The emergence of privacy computing technology provides a solution for the secure circulation of data. It can ensure that our data is "available but invisible" and will not be directly observed while being used externally. Among them, federated learning is one of the key technologies of privacy computing. It is a distributed learning paradigm that realizes the basic idea of ​​"data is not fixed, model is moving" by jointly training models with multiple data holders, i.e., participants. Furthermore, by introducing the federated learning framework in the training of graph neural networks, it is possible to collaboratively train graph models under the premise of distributed storage of graph data. At the same time, federated learning can also make full use of the computing power of terminal devices and reduce the computing cost and data storage requirements of the server. However, in real life, the local data distribution of different participants is usually inconsistent, that is, there is a problem of data heterogeneity, which makes the global shared model obtained by federated average training not applicable to the local data of the participants. In addition, distributedly stored graph data has natural heterogeneity in topological structure and feature information. Therefore, the problem of data heterogeneity is more prominent in the scenario of federated graph learning, resulting in serious model differences during federated learning training, thereby reducing the effectiveness of the global model for local data.

[0004] In order to solve the problem of low global model utility caused by data heterogeneity in federated graph learning scenarios, existing solutions are divided into two categories: one is the global model method, that is, training a globally shared graph model based on the graph data of all participants. Common strategies are to reconstruct the loss function, redesign the weighted aggregation coefficient and the model interpolation method; the other is the personalized model method, that is, training a personalized model for each participant. The common strategy is to cluster participants, that is, participants with similar data distribution are divided into the same group to share the model. However, although the existing solutions can alleviate the above problems to a certain extent, most of the research on the global model method lacks comprehensive consideration of the topological structure of the graph data and the node feature information, and the research scheme of the personalized model also has similar limitations. Most of them are based on clustering methods and the ideas are relatively single. According to the particularity of the training data type in the federated graph learning scenario, it is necessary to consider more the structural characteristics of the local graph data of the participants to build a customized local model, so as to better adapt to the local data. Summary of the invention

[0005] In view of the above problems existing in the prior art, the present invention is proposed.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a personalized federated graph learning method based on deep graph similarity learning, comprising: collecting local map data, and obtaining an unsupervised graph encoder through horizontal federated average learning; inputting the local map data into the unsupervised graph encoder to calculate a dense vector; inputting the dense vector into a central server to calculate the similarity and construct a personalized model; and performing federated iterative training on the participants according to the personalized model.

[0007] As a preferred solution of the personalized federated graph learning method based on deep graph similarity learning described in the present invention, wherein: the collection of local map data includes: the participant C i The features and adjacency information of the local graph nodes are extracted and integrated, and the data is represented as:

[0008] G i =[V i ,E i ]

[0009] Among them, G i Represented as the local data of the participant, V i Represented as a node set, E i Represented as an edge set.

[0010] As a preferred solution of the personalized federated graph learning method based on deep graph similarity learning described in the present invention, the unsupervised graph encoder includes that the participating party trains the unsupervised encoder based on local data graph data through federated average learning to obtain a graph encoder containing global information.

[0011] As a preferred solution of the personalized federated graph learning method based on deep graph similarity learning described in the present invention, wherein: the calculation of dense vectors includes that each participant inputs the local graph data into the unsupervised graph encoder to calculate the dense vector, which is expressed as,

[0012] μ i =f(G i )

[0013] Among them, μ i Represented as a dense vector, G i is represented as the local data of the participants, and f is represented as an unsupervised graph encoder.

[0014] As a preferred solution of the personalized federated graph learning method based on deep graph similarity learning described in the present invention, wherein: the inputting of dense vectors into the central server to calculate the similarity includes calculating the similarity expressed as,

[0015]

[0016] Among them, S(i,j) is the participant C i and Party C j The similarity between .

[0017] As a preferred solution of the personalized federated graph learning method based on deep graph similarity learning described in the present invention, the personalized model includes: the central server selects the set of participants participating in the training in this round based on the strategy of uniform label distribution in each iteration, and the server calculates the weighted aggregation coefficient of the personalized model of the participants, which is expressed as,

[0018]

[0019] Among them, α ij is the weighted aggregation coefficient, S(i,j) is the number of participants C i and Party C j The similarity between

[0020] The personalized model parameters of the participants are aggregated and calculated as follows:

[0021]

[0022] Among them, θ i For participant C i Personalized model, θ j For participant C j Personalized model;

[0023] As a preferred solution of the personalized federated graph learning method based on deep graph similarity learning described in the present invention, wherein: the federated iterative training of the participants according to the personalized model includes calculating the similarity of the participants and performing aggregation calculations on the personalized models of the participants as a training round, and the server detects the learning termination conditions at the end of each round. If the termination conditions are met, the server initiates a learning termination broadcast to the participants and determines that the training is terminated. If the termination conditions are not met, the iterative training continues.

[0024] As a preferred solution of the personalized federated graph learning system based on deep graph similarity learning described in the present invention, it includes: a data collection module, a vector calculation module, a model building module, and an iterative training module;

[0025] The data collection module collects local map data and obtains an unsupervised map encoder through horizontal federated average learning;

[0026] The vector calculation module inputs the local graph data into the unsupervised graph encoder to calculate the dense vector;

[0027] The model building module inputs the dense vector into the central server to calculate the similarity and build a personalized model;

[0028] The iterative training module performs federated iterative training on the participants according to the personalized model.

[0029] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of any one of the methods described in the personalized federated graph learning method based on deep graph similarity learning are implemented.

[0030] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of any one of the methods described in the personalized federated graph learning method based on deep graph similarity learning are implemented.

[0031] Beneficial effects of the present invention: The personalized federated graph learning method based on deep graph similarity learning disclosed in the present invention is intended to provide personalized models for participants with data heterogeneity so that they can adapt to local graph data. This framework consists of two core parts: a personalized parameter aggregation mechanism based on deep graph similarity learning and a participant selection mechanism based on uniform label distribution. The former generates a low-dimensional dense vector based on the local graph data of the participants by collaboratively training an unsupervised graph encoder. The central server calculates the similarity between the two participants based on the vectors sent by each participant. Through a normalized weighted aggregation method, the personalized model parameters are calculated for each federated participant and distributed and updated, so that the model actually obtained by the participant can be adapted to the local data. The latter optimizes the participant selection process in the federated learning training process. Under the premise of retaining randomness, the set of participants participating in this round of training is selected based on the number of participants' samples and label distribution, so that the data actually participating in the training is distributed as evenly as possible, and the representation of information is more comprehensive, thereby improving the universality and generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0033] Figure 1 A schematic flowchart of a personalized federated graph learning method based on deep graph similarity learning is provided for one embodiment of the present invention.

[0034] Figure 2 A process framework diagram of a personalized federated graph learning method based on deep graph similarity learning provided for an embodiment of the present invention.

[0035] Figure 3 A schematic diagram of federated graph neural network learning for a personalized federated graph learning method based on deep graph similarity learning provided for an embodiment of the present invention.

[0036] Figure 4 An unsupervised encoder flow chart of a personalized federated graph learning method based on deep graph similarity learning provided for one embodiment of the present invention.

[0037] Figure 5 A random walk sampling graph based on earliest access time of a personalized federated graph learning method based on deep graph similarity learning provided by an embodiment of the present invention.

[0038] Figure 6A participant selection strategy diagram based on uniform label distribution of a personalized federated graph learning method based on deep graph similarity learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0040] Example 1, reference Figure 1-Figure 3 , which is the first embodiment of the present invention, and provides a personalized federated graph learning method based on deep graph similarity learning, including:

[0041] S1: Collect local graph data and obtain an unsupervised graph encoder through horizontal federated averaging learning.

[0042] Collecting local map data includes: i The features and adjacency information of the local graph nodes are extracted and integrated, and the data is represented as:

[0043] G i =[V i ,E i ]

[0044] Among them, G i Represented as the local data of the participant, V i Represented as a node set, E i Represented as an edge set.

[0045] It should be noted that when participants in horizontal federated graph learning prepare local graph datasets, they need to formulate detailed strategies based on the characteristics of the data and the needs of federated learning. First, participants need to comprehensively organize and analyze their locally stored graph data, including extracting and cleaning the information of graph node features and edges to ensure the integrity and consistency of the data. At the same time, based on the distribution characteristics of the data and the training requirements of the subsequent federated learning model, the training set and test set should be reasonably divided to provide support for the two stages of federated learning. When dividing the data set, a variety of data partitioning strategies can be considered, such as random partitioning, time-based partitioning, or hierarchical partitioning by node attributes to ensure that the training set and test set can cover the main characteristics of the local data. At the same time, in order to further optimize the training effect of the model, participants can statistically analyze the characteristic distribution of the data in advance, and record the relevant statistical information as metadata to provide a reference for the subsequent federated learning stage. In addition, participants can also perform subgraph sampling and encoding operations on the data set to reduce computational complexity and improve data processing efficiency during the actual training process, such as Figure 1 S1 and Figure 2 shown.

[0046] Furthermore, the unsupervised graph encoder includes that the participants train the unsupervised encoder based on local data graph data through federated average learning to obtain a graph encoder containing global information.

[0047] It should be noted that the unsupervised graph encoder is a technology that can embed graph structures without labeled data. Under the framework of federated learning, each participant uses the locally stored graph data to collaboratively train the unsupervised graph encoder through horizontal federated average learning. During the training process, each participant first independently initializes the encoder parameters based on its local data, and then performs several rounds of local training to update the model parameters. After the participant completes the local update, the model parameters are sent to the central server, which performs weighted aggregation on the received parameters and returns the aggregated parameters to each participant to complete the update. Through this cyclic process, each participant gradually builds a high-quality graph encoder containing global information. The finally trained unsupervised graph encoder can effectively integrate the local graph data characteristics of each participant while maintaining data privacy. This method not only solves the problem of graph data islands caused by distributed storage, but also balances the heterogeneity between participants through the aggregation strategy of the central server, so that the encoder can be applied to a variety of data distribution scenarios. Based on this encoder, each participant can generate a compact dense vector representation, which lays the foundation for further personalized model training and data similarity analysis. This approach is particularly suitable for achieving efficient distributed learning and collaboration in privacy-sensitive environments.

[0048] S2: Input the local graph data into the unsupervised graph encoder to calculate the dense vector.

[0049] It should be noted that the calculation of the dense vector includes that each participant inputs the local graph data into the unsupervised graph encoder to calculate the dense vector, which is expressed as,

[0050] μ i =f(G i )

[0051] Among them, μ i Represented as a dense vector, G i is represented as the local data of the participants, and f is represented as an unsupervised graph encoder.

[0052] Furthermore, in the process of calculating dense vectors, each participant inputs the local graph data into the unsupervised graph encoder for processing. The core function of the unsupervised graph encoder is to extract the key features of the nodes and their neighborhoods in the graph by encoding and decoding the graph data, and convert these features into low-dimensional dense vector representations. Each participant preprocesses its local graph data, which includes extracting the feature information of the nodes and the adjacency matrix from the original graph data to ensure that the data can be correctly understood by the encoder. In the input stage, the encoder will perform multiple subgraph sampling on these preprocessed graph data to obtain a series of representative subgraphs. Through the sampling method based on random walks, the encoder can effectively capture the local relations in the graph structure. The encoder uses automatic encoding technology to process the sampled subgraphs. The encoder maps the feature representation of the node to a low-dimensional space and generates compact dense vectors, which can effectively represent the main information of the nodes and their surrounding structures in the graph. In addition, in order to ensure the quality of the encoding, the encoder will perform reverse verification through the decoding stage, that is, re-decoding the generated dense vector into the original graph structure and calculating the reconstruction error. This self-supervised training method ensures that the dense vector is not only compact but also has high fidelity.

[0053] After the above process, each participant generates a dense vector representation of its local graph data. These dense vectors are not only an abstract representation of the graph data, but also provide a unified processing interface for subsequent steps, facilitating similarity calculation and aggregation optimization of personalized model parameters. These dense vectors will be transmitted to the central server for further calculation of graph similarities between participants, thus laying the foundation for the personalized federated learning process.

[0054] S3: Input the dense vector into the central server to calculate the similarity and build a personalized model.

[0055] It should be noted that inputting the dense vector into the central server to calculate the similarity includes calculating the similarity expressed as,

[0056]

[0057] Among them, S(i,j) is the participant C i and Party C j The similarity between .

[0058] After receiving these vectors, the central server calculates the graph data similarity between different participants and compares the dense vectors of each pair of participants. This comparison can not only effectively capture the topological structure characteristics and node feature distribution of the local graph data, but also reflect the potential degree of association between the participants. In this way, the degree of similarity can provide a basis for the subsequent weighted aggregation of personalized model parameters.

[0059] Furthermore, the personalized model includes that the central server selects the set of participants participating in the training in each round based on the strategy of uniform label distribution in each iteration, and the server calculates the weighted aggregation coefficient of the personalized model of the participants, which is expressed as,

[0060]

[0061] Among them, α ij is the weighted aggregation coefficient, S(i,j) is the number of participants C i and Party C j The similarity between

[0062] The personalized model parameters of the participants are aggregated and calculated as follows:

[0063]

[0064] Among them, θ i For participant C i Personalized model, θ j For participant C j Personalized model.

[0065] Each participant initializes the local initial graph neural network model parameters. For the graph neural network model, Figure 3 As shown in the figure, the main learning parameters are the learnable parameters in the AGGREGATE and UPDATE functions. After the model training iteration begins, for the tth iteration, the central server selects K participants to participate in this round of training based on the uniform label distribution strategy. The selected K participants perform local model training and gradient descent based on local data. The participants send the model parameters obtained from the local training to the central server for personalized parameter weighted aggregation based on graph similarity. i The personalized model parameters are aggregated and calculated with the participating party C i The higher the similarity of the local graph of the participant, the higher the weight of the model parameter.i The lower the local graph similarity of a participant, the lower the weight value of its model parameter.

[0066] S4: Federate and iteratively train participants based on the personalized model.

[0067] It should be noted that the federated iterative training of the participants according to the personalized model includes calculating the similarity of the participants and performing aggregation calculations on the personalized models of the participants as a training round. The server detects the learning termination conditions at the end of each round. If the termination conditions are met, the server initiates a learning termination broadcast to the participants and determines that the training is over. If the termination conditions are not met, the iterative training continues.

[0068] Furthermore, each training round consists of multiple key steps. The central server first receives the local dense vector representations of all participants. These vectors are usually low-dimensional feature representations extracted by the participants from their local map data through unsupervised encoders. Based on the predetermined similarity measurement method, the server calculates the similarity score between each two participants and constructs a complete similarity matrix. The server uses the similarity matrix to weighted aggregate the local model parameters of the participants. The weights of the model parameters of other participants with more similar data distribution to a certain participant are assigned higher values ​​during aggregation; while the parameter influence of participants with lower similarity is weakened. This personalized weighting method ensures that the final generated model is more in line with the local data characteristics of the participants. The aggregated personalized model parameters will be distributed to the corresponding participants. The participants use these parameters to continue the next round of training and optimization on the local data. The server detects the learning termination conditions at the end of each round. If the termination conditions are met, the server initiates a learning termination broadcast to the participants to determine that the training is over. If the termination conditions are not met, iterative training will continue.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0070] Embodiment 2, the second embodiment of the present invention, is different from the previous embodiment in that:

[0071] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0072] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0073] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0074] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0075] Embodiment 3 is the third embodiment of the present invention, which provides a personalized federated graph learning system based on deep graph similarity learning, characterized by: comprising a data collection module, a vector calculation module, a model building module, and an iterative training module;

[0076] The data collection module collects local map data and obtains an unsupervised map encoder through horizontal federated average learning;

[0077] The vector calculation module inputs the local graph data into the unsupervised graph encoder to calculate the dense vector;

[0078] The model building module inputs the dense vector into the central server to calculate the similarity and build a personalized model;

[0079] The iterative training module performs federated iterative training on the participants according to the personalized model.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0081] Example 4, reference Figure 4-Figure 6 , which is an embodiment of the present invention, provides a personalized federated graph learning method based on deep graph similarity learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0082] The construction process of the unsupervised encoder includes three main modules: sampling, encoding, and embedding representation distribution, such as Figure 4-Figure 5 shown.

[0083] The sampling includes performing multiple subgraph sampling by random walk based on the earliest access time to obtain several subgraphs, and then concatenating the node feature vectors in each subgraph to represent the input feature of the subgraph as {X1,X2,...,X n The purpose of this stage is to find a class of subgraphs that can be effectively encoded and decoded so that the reconstruction error can be evaluated in an easy-to-handle way. The sampling process of k nodes is as follows:

[0084] Randomly extract a node v from the graph data (0) As the starting node, and set its earliest access time to 0, start the wandering process;

[0085] Select the next hop node: Assume v (p) is the node visited at timestamp, from v (p) A node v is randomly selected from the one-hop neighborhood or first-order neighbor node of (p+1) as the node to be visited at timestamp p+1. If v (p+1) If it is a node that has not been visited in the previous time, its earliest visit time is set to p+1; otherwise, the earliest visit time of the node remains unchanged;

[0086] When the timestamp reaches k, the sampling process terminates.

[0087] The encoding includes using the autoencoder to encode the subgraphs obtained in the previous step to obtain dense low-dimensional vectors. For each subgraph X i , the autoencoder performs the following encoding and decoding process:

[0088] z i =f(X i ,θ e )

[0089]

[0090] Among them, z i Indicates that the input subgraph contains the feature X of the node i The low-dimensional dense embedding representation of e , the decoder function parameter is θ d , the following reconstruction loss function is used to train the above encoder and decoder:

[0091]

[0092] The purpose of this stage is to find an autoencoder to obtain a higher quality vector representation of the subgraph nodes, so as to better reconstruct the original subgraph based on the vector representation. After the above autoencoder is trained, the embedded vector representation {z1,z2,...,z n}, showing the distribution of each sub-graph.

[0093] Embedding the representation distribution involves evaluating the distance between the subgraph distributions of the two original input graphs G and H using the maximum mean squared deviation (MMD). For two graphs, their distance in the representation space should be similar to the distance between their subgraph vector distributions. Suppose k subgraphs are sampled from the input graph G and encoded as {g1,g2,...,g k}, the k subgraphs of the input graph H are encoded as {h1,h2,...,h k}, then the MMD distance between graph G and graph H is:

[0094]

[0095] in, and It is the empirical kernel embedding representation of two distributions, which is usually defined by the identity kernel. The implementation is very simple. No transformation is performed on the input, and the input is directly mapped to the identity, that is, the input is directly used as the output. Then and It can be expressed as:

[0096]

[0097] The output representation of a subgraph is obtained by averaging and aggregating its subgraph representations. In this way, the similarity of any two input graphs G and H can be evaluated, and MMD is used as a similarity metric. The smaller the distance between the two graphs, the more similar they are, so that personalized weight parameters can be further calculated for each participant.

[0098] The basic idea of ​​the participant selection strategy based on uniform label distribution is that when the central server selects participants, it will select a group of participants to participate in this round of training based on the label distribution of each participant's local data, such as Figure 6 The distribution of labels of these participants needs to be as evenly distributed as possible after overall aggregation, that is, the number of labels of each category should be as close as possible, so that the representation of information is more comprehensive and the generalization of the model is improved. The details are as follows:

[0099] Participants need to collect statistics on the label distribution of local data in advance and send the label distribution to the central server;

[0100] Before each round of federated training begins, the central server selects participants. If the number of participants actually participating in each round of training is set to M, the central server needs to randomly select K times more participants [C1, C2, ..., C K·M ] to ensure that the selection of participants has a certain degree of randomness, so that each participant has the possibility to participate in model training;

[0101] After that, the central server needs to K·M ] and then select M participants so that the comprehensive distribution of labels of the participants finally participating in the training is close to a uniform distribution.

[0102] The comprehensive distribution of the labels of the participants can be achieved by a simple permutation and combination method, that is, the label distribution of each possible combination of participants is weighted and summed based on the number of samples to obtain a label distribution after the combination of the participants, and then the difference or distance is calculated with the uniform distribution set by the central server. The difference calculation here can be achieved by calculating the Wasserstein distance between the two distributions (a distance metric used to measure the degree of difference between two probability distributions).

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A personalized federated graph learning method based on deep graph similarity learning, characterized by: include: Collect local graph data and obtain an unsupervised graph encoder through horizontal federated averaging learning; Input the local graph data into the unsupervised graph encoder to calculate the dense vector; The dense vector is input into the central server to calculate the similarity and build a personalized model; Federated iterative training of participants based on personalized models.

2. The personalized federated graph learning method based on deep graph similarity learning according to claim 1, characterized in that: The collection of local map data includes: i The features and adjacency information of the local graph nodes are extracted and integrated, and the data is represented as: G i =[V i ,E i ] Among them, G i Represented as the local data of the participant, V i Represented as a node set, E i Represented as an edge set.

3. The personalized federated graph learning method based on deep graph similarity learning as claimed in claim 2, characterized in that: The unsupervised graph encoder includes that the participants train the unsupervised encoder based on local data graph data through federated average learning to obtain a graph encoder containing global information.

4. The personalized federated graph learning method based on deep graph similarity learning as claimed in claim 3, characterized in that: The calculation of the dense vector includes that each participant inputs the local graph data into the unsupervised graph encoder to calculate the dense vector, which is expressed as: m i =f(G i ) Among them, μ i Represented as a dense vector, G i is represented as the local data of the participants, and f is represented as an unsupervised graph encoder.

5. The personalized federated graph learning method based on deep graph similarity learning as claimed in claim 4, characterized in that: The inputting the dense vector into the central server to calculate the similarity includes calculating the similarity expressed as: Among them, S(i,j) is the participant C i and Party C j The similarity between .

6. The personalized federated graph learning method based on deep graph similarity learning as claimed in claim 5, characterized in that: The personalized model includes that the central server selects the set of participants participating in the training in this round based on the strategy of uniform label distribution in each iteration, and the server calculates the weighted aggregation coefficient of the personalized model of the participants, which is expressed as: Among them, α ij is the weighted aggregation coefficient, S(i,j) is the number of participants C i and Party C j The similarity between The personalized model parameters of the participants are aggregated and calculated as follows: Among them, θ i For participant C i Personalized model, θ j For participant C j Personalized model.

7. The personalized federated graph learning method based on deep graph similarity learning according to claim 6, characterized in that: The federated iterative training of the participants according to the personalized model includes calculating the similarity of the participants and performing aggregation calculation on the personalized models of the participants as a training round. The server detects the learning termination condition at the end of each round. If the termination condition is met, the server initiates a learning termination broadcast to the participants and determines that the training is over. If the termination condition is not met, the iterative training continues.

8. A personalized federated graph learning system based on deep graph similarity learning according to any one of claims 1 to 7, characterized in that: Including data collection module, vector calculation module, model building module, iterative training module; The data collection module collects local map data and obtains an unsupervised map encoder through horizontal federated average learning; The vector calculation module inputs the local graph data into the unsupervised graph encoder to calculate the dense vector; The model building module inputs the dense vector into the central server to calculate the similarity and build a personalized model; The iterative training module performs federated iterative training on the participants according to the personalized model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images