Federal recommendation system model aggregation strategy research method based on graph neural network

Abstract

The invention relates to the technical field of neural networks, and discloses a federal recommendation system model aggregation strategy research method based on a graph neural network, and the method comprises the steps: a data preprocessing stage; a local model training stage; a model parameter encryption and transmission stage; in the federal aggregation stage, decryption operation is carried out, and federal average or other advanced federal learning aggregation algorithms are applied; in the global model updating and distributing stage, the aggregated global model parameters are re-encrypted and distributed back to each client; and an iterative optimization and evaluation stage: carrying out iterative optimization. According to the method, encryption processing is adopted in the model parameter transmission process, so that the privacy and security of data in the transmission and aggregation stages are ensured, and the privacy risk caused by leakage of local model parameters is effectively avoided; a differential privacy technology and a block chain mechanism are further introduced, noise is added in model parameters, distributed storage and verification are realized, and the privacy protection capability of the system is further improved.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a research method for a federated recommendation system model aggregation strategy based on a graph neural network. Background Art

[0002] With the rapid development of big data and artificial intelligence technology, recommendation systems have become an effective tool to solve the problem of "information overload". Recommendation systems analyze user behavior, interests and needs, mine information or services that users may be interested in from massive data, and push them to users in an intelligent and personalized way. In recent years, graph neural networks, as a deep learning method that can directly learn graph structured data, have attracted much attention because of their ability to capture complex high-order relationships between users and items. Among them, Federated Learning (FL) has emerged as a privacy-preserving machine learning technology.

[0003] However, in actual use, the application of federated recommendation systems still faces many challenges. First, local models may leak privacy during transmission and aggregation, causing corresponding information security risks. Second, since the amount of local "user-item" interaction data is usually small, it is difficult to train an accurate model. In addition, local user data usually only includes first-order "user-item" interaction information and cannot be extended to high-order information, which limits the recommendation effect of the model. Based on this, the present invention designs a research method for model aggregation strategy of federated recommendation systems based on graph neural networks to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a research method for a federated recommendation system model aggregation strategy based on graph neural network, which solves the problems in the background technology.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: A research method for a model aggregation strategy of a federated recommendation system based on a graph neural network comprises: step (1), a data preprocessing stage, for collecting and cleaning data and extracting features from the cleaned data; step (2), a local model training stage, for performing local training using a graph neural network model and optimizing model parameters to predict user preferences; step (3), a model parameter encryption and transmission stage, for encrypting the GNN model parameters trained on each client; step (4), a federated aggregation stage, for receiving the encrypted model parameters from each client, performing a decryption operation, and applying a federated average or other advanced federated learning aggregation algorithm; step (5), a global model update and distribution stage, for re-encrypting the aggregated global model parameters and distributing them back to each client; step (6), an iterative optimization and evaluation stage, for performing iterative optimization until the preset training rounds or the model performance improvement is no longer significant.

[0006] Preferably, the step (1) comprises the following steps: a. Collect user behavior data on multiple distributed clients, including but not limited to user click, purchase, and evaluation history records; b. Clean the collected data to remove invalid, duplicate, or abnormal data to ensure data quality; c. Perform feature extraction on the cleaned data and construct a user-item interaction graph, where nodes represent users or items and edges represent the interaction between users and items.

[0007] Preferably, the step (2) comprises the following steps: a. On each client, based on the constructed user-item interaction graph, a graph neural network (GNN) model is used for local training to learn the potential representations of users and items. b. During the local training process, a graph convolutional network (GCN), graph attention network (GAT) or other graph neural network variants are used to combine user historical behavior and item attributes to optimize model parameters to predict user preferences.

[0008] Preferably, the step (3) comprises the following steps: a. Encrypt the GNN model parameters trained on each client to ensure data privacy and security during transmission; b. Transmit the encrypted model parameters to the central server for subsequent model aggregation; The encryption process in step (3) adopts homomorphic encryption, differential privacy or secure multi-party computing technology to ensure the security of data transmission.

[0009] Preferably, step (4) comprises the following steps: a. the central server receives the encrypted model parameters from each client and performs a decryption operation; b. applies federated averaging or other advanced federated learning aggregation algorithms to aggregate the decrypted model parameters to form a global model; c. optionally, introduces model regularization or weight decay technology to improve the generalization ability of the global model; The federated averaging algorithm in step (4) further considers the amount of data and model training quality of each client, and adopts a weighted averaging strategy for model aggregation.

[0010] Preferably, the step (5) comprises the following steps: a. re-encrypting the aggregated global model parameters and distributing them back to each client; b. each client receives the updated global model parameters, replaces the local model, and prepares for the next round of training; In step (5), when the global model is updated, the stability and convergence speed of the model are also considered, and the learning rate is dynamically adjusted or other optimization strategies are adopted.

[0011] Preferably, the step (6) comprises the following steps: a. Repeat steps (2) to (5) until the preset training rounds are reached or the model performance is no longer significantly improved; b. After each round of iteration, perform a performance evaluation on the global model, including common indicators of recommendation systems such as accuracy, recall, and F1 score.

[0012] A research method for a federated recommendation system model aggregation strategy based on graph neural network also includes the following optional steps: Step (i). Model personalization adjustment stage: After the global model is distributed to each client, the global model is fine-tuned or customized according to the specific needs of each client, user group characteristics or local data set characteristics to further improve the recommendation effect and user satisfaction; this step involves adjusting the model's learning rate, regularization parameters, or introducing additional local features for training; Step (ii). Communication efficiency optimization stage. Considering the communication overhead in federated learning, this method can further optimize the transmission process of model parameters; use parameter compression technology to reduce the amount of data transmitted, or use model distillation technology to extract key information to reduce communication costs; design an efficient model parameter update strategy, through asynchronous update or gradient-based update, to reduce communication rounds and waiting time; Step (iii). In the privacy protection enhancement stage, based on the encryption and transmission of model parameters, a higher level of privacy protection mechanism is further introduced; by using differential privacy technology to add noise to the model parameters, it is ensured that even if the model parameters are leaked, the specific user data cannot be inferred; at the same time, a distributed storage and verification mechanism based on blockchain can be explored to enhance the transparency and security of the model parameter transmission and aggregation process; Step (iv). In the stage of improving the robustness of the federated recommendation system, this method can introduce adversarial training or robustness optimization technology to improve the robustness of the federated recommendation system. By introducing adversarial samples or noise, the training model can still maintain good performance when facing data disturbances or malicious attacks. In addition, anomaly detection and fault-tolerant mechanisms are designed to deal with potential problems such as client failures or data pollution.

[0013] Preferably, in step (i), based on the global model, each client can perform fine-tuning according to the uniqueness of local data; this includes adjusting the model architecture, optimizing the objective function, or introducing additional regularization terms to adapt to different user groups or application scenarios.

[0014] Preferably, in step (ii), before the model parameters are transmitted, a parameter compression algorithm is used to preprocess the model parameters to reduce the amount of data transmitted, which includes methods such as quantizing parameter values, sparse representation, or using low-rank decomposition; and by designing efficient communication protocols and scheduling strategies to minimize communication delays and bandwidth occupancy.

[0015] Preferably, in step (iii), on the basis of encrypting the model parameters, differential privacy technology is further used to protect user privacy; by adding an appropriate amount of noise to the model parameters, it is ensured that the data of a single user cannot be accurately inferred during the model training process; blockchain technology is introduced to realize the distributed storage and verification of model parameters; and the security and credibility of the model parameters are ensured by the immutability and decentralization characteristics of blockchain.

[0016] Preferably, in step (iv), adversarial training technology is used to train the model by introducing adversarial samples or noise to improve the robustness of the model to input data disturbances; anomaly detection and fault tolerance mechanisms are designed to detect and respond to potential problems such as client failures, data contamination or malicious attacks.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are: 1. In the present invention, encryption processing is used in the process of model parameter transmission to ensure the privacy and security of data in the transmission and aggregation stages, and effectively avoid the privacy risks caused by the leakage of local model parameters. Differential privacy technology and blockchain mechanism are further introduced to add noise to the model parameters, and distributed storage and verification are realized. Even if the model parameters are leaked, the specific user data cannot be accurately inferred, which further enhances the privacy protection capability of the system.

[0018] 2. In the present invention, a graph neural network (GNN) model is used for local training, which can learn the potential representations of users and items, and optimize the model parameters in combination with the user's historical behavior and item attributes, thereby improving the model's prediction accuracy for user preferences; the decrypted model parameters are aggregated through federated averaging or other advanced federated learning aggregation algorithms to form a global model, and at the same time, model regularization or weight decay technology is introduced to improve the generalization ability of the global model.

[0019] 3. In the present invention, by constructing a user-item interaction graph and expanding the first-order interaction information to high-order information, the input data of the model is enriched and the recommendation effect of the model is improved. After the global model is distributed to each client, fine-tuning or customized adjustments are performed according to the specific needs of each client, user group characteristics or the characteristics of the local data set, thereby improving the recommendation effect and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a system flow chart of the research method of the present invention; Figure 2 For the present invention Figure 1 Detailed diagram of the process; Figure 3 Optional step diagram of the research method of the present invention; Figure 4 For the present invention Figure 3 Detailed diagram of the process. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] See also Figure 1-Figure 2In an embodiment of the present invention, a research method for a model aggregation strategy of a federated recommendation system based on a graph neural network includes: step (1), a data preprocessing stage, which is used to collect and clean data and extract features from the cleaned data; step (2), a local model training stage, which uses a graph neural network model to perform local training and optimize model parameters to predict user preferences; step (3), a model parameter encryption and transmission stage, which encrypts the GNN model parameters trained on each client; step (4), a federated aggregation stage, which receives the encrypted model parameters from each client, performs a decryption operation, and applies a federated average or other advanced federated learning aggregation algorithm; step (5), a global model update and distribution stage, which re-encrypts the aggregated global model parameters and distributes them back to each client; step (6), an iterative optimization and evaluation stage, which performs iterative optimization until the preset training rounds or the model performance improvement is no longer significant.

[0023] Step (1) includes the following steps: a. Collecting user behavior data on multiple distributed clients, including but not limited to user click, purchase, evaluation and other historical records; b. Cleaning the collected data to remove invalid, duplicate or abnormal data to ensure data quality; c. Extracting features from the cleaned data and constructing a user-item interaction graph, where nodes represent users or items and edges represent the interaction between users and items.

[0024] Step (2) includes the following steps: a. On each client, based on the constructed user-item interaction graph, a graph neural network (GNN) model is used for local training to learn the potential representations of users and items; b. During the local training process, a graph convolutional network (GCN), a graph attention network (GAT) or other graph neural network variants is used to combine user historical behavior and item attributes to optimize model parameters to predict user preferences.

[0025] Step (3) includes the following steps: a. Encrypting the GNN model parameters trained on each client to ensure data privacy and security during transmission; b. Transmitting the encrypted model parameters to the central server for subsequent model aggregation; The encryption process in step (3) uses homomorphic encryption, differential privacy or secure multi-party computing technology to ensure the security of data transmission.

[0026] Step (4) includes the following steps: a. The central server receives the encrypted model parameters from each client and performs a decryption operation; b. Federated Averaging or other advanced federated learning aggregation algorithms are applied to aggregate the decrypted model parameters to form a global model; c. Optionally, model regularization or weight decay technology is introduced to improve the generalization ability of the global model; The federated averaging algorithm in step (4) further considers the data volume and model training quality of each client and adopts a weighted average strategy for model aggregation.

[0027] Step (5) includes the following steps: a. re-encrypting the aggregated global model parameters and distributing them back to each client; b. each client receives the updated global model parameters, replaces the local model, and prepares for the next round of training; in step (5), when the global model is updated, the stability and convergence speed of the model are also considered, and the learning rate is dynamically adjusted or other optimization strategies are adopted.

[0028] Step (6) includes the following steps: a. Repeat steps (2) to (5) until the preset training rounds are reached or the model performance improvement is no longer significant; b. After each round of iteration, the performance of the global model is evaluated, including common indicators of recommendation systems such as accuracy, recall rate, and F1 score.

[0029] The working principle of the embodiment of the present invention is as follows: In the data preprocessing stage (step 1), the present invention first collects user behavior data through multiple distributed clients, including user click, purchase, evaluation and other historical records. These data are the basis for training the recommendation system model. Next, the collected data is cleaned to remove invalid, duplicate or abnormal data to ensure data quality. Finally, feature extraction is performed on the cleaned data to construct a user-item interaction graph. This step provides effective data input for subsequent graph neural network model training.

[0030] Entering the local model training phase (step 2), the present invention uses a graph neural network (GNN) model for local training based on the constructed user-item interaction graph on each client. Graph neural networks can capture the complex relationship between users and items and learn the potential representations of users and items. During the training process, the present invention uses a graph convolutional network (GCN), a graph attention network (GAT) or other graph neural network variants, combined with user historical behavior and item attributes, to optimize model parameters to predict user preferences. This step enables each client to train a preliminary recommendation model based on local data.

[0031] In the model parameter encryption and transmission stage (step 3), the present invention encrypts the GNN model parameters trained on each client to ensure data privacy and security during the transmission process. The encryption process uses homomorphic encryption, differential privacy or secure multi-party computing technology, which can ensure the security of data transmission and prevent data leakage. The encrypted model parameters are transmitted to the central server for subsequent model aggregation.

[0032] The federated aggregation stage (step 4) is the core of the method of the present invention. The central server receives the encrypted model parameters from each client and performs a decryption operation. Then, the decrypted model parameters are aggregated by applying Federated Averaging or other advanced federated learning aggregation algorithms to form a global model. The Federated Averaging algorithm takes into account the data volume and model training quality of each client, and adopts a weighted average strategy for model aggregation to ensure the accuracy and robustness of the global model. In addition, in order to improve the generalization ability of the global model, the present invention also introduces model regularization or weight decay technology.

[0033] In the global model update and distribution phase (step 5), the present invention re-encrypts the aggregated global model parameters and distributes them back to each client. Each client receives the updated global model parameters, replaces the local model, and prepares for the next round of training. This step implements the update and distribution of the global model parameters, so that each client can train based on the latest global model. At the same time, in order to improve the stability and convergence speed of the model, the present invention also considers dynamically adjusting the learning rate or adopting other optimization strategies.

[0034] Finally, in the iterative optimization and evaluation phase (step 6), the present invention repeats steps 2 to 5 until the preset training rounds are reached or the model performance improvement is no longer significant. After each round of iteration, the global model is evaluated for performance, including common indicators of recommendation systems such as accuracy, recall, and F1 score. This step ensures continuous optimization and performance improvement of the model during the training process.

[0035] See also Figure 1-Figure 4 In an embodiment of the present invention, a method for studying aggregation strategies of a federated recommendation system model based on a graph neural network further includes the following optional steps: Step (i). Model personalization adjustment stage: After the global model is distributed to each client, the global model is fine-tuned or customized according to the specific needs of each client, user group characteristics or local data set characteristics to further improve the recommendation effect and user satisfaction; this step involves adjusting the model's learning rate, regularization parameters, or introducing additional local features for training; Step (ii). Communication efficiency optimization stage. Considering the communication overhead in federated learning, this method can further optimize the transmission process of model parameters; use parameter compression technology to reduce the amount of data transmitted, or use model distillation technology to extract key information to reduce communication costs; design an efficient model parameter update strategy, through asynchronous update or gradient-based update, to reduce communication rounds and waiting time; Step (iii). In the privacy protection enhancement stage, based on the encryption and transmission of model parameters, a higher level of privacy protection mechanism is further introduced; by using differential privacy technology to add noise to the model parameters, it is ensured that even if the model parameters are leaked, the specific user data cannot be inferred; at the same time, a distributed storage and verification mechanism based on blockchain can be explored to enhance the transparency and security of the model parameter transmission and aggregation process; Step (iv). In the stage of improving the robustness of the federated recommendation system, this method can introduce adversarial training or robustness optimization technology to improve the robustness of the federated recommendation system. By introducing adversarial samples or noise, the training model can still maintain good performance when facing data disturbances or malicious attacks. In addition, anomaly detection and fault-tolerant mechanisms are designed to deal with potential problems such as client failures or data pollution.

[0036] The working principle of the embodiment of the present invention is: after the global model is distributed to each client, the present invention first enters the model personalization adjustment stage. The core goal of this stage is to fine-tune or customize the global model according to the specific needs of each client, the characteristics of the user group or the characteristics of the local data set, so as to further improve the recommendation effect and user satisfaction. To achieve this goal, the present invention allows each client to adjust the model's learning rate, regularization parameters and other hyperparameters according to local actual conditions after receiving the global model, or introduce additional local features for training. These adjustment operations can make full use of the local data advantages of each client to make the model more in line with the preferences of a specific user group, thereby significantly improving the recommendation accuracy and user experience.

[0037] Taking into account the communication overhead problem in federated learning, the present invention further optimizes the model parameter transmission process. On the one hand, parameter compression technology is used to reduce the amount of data transmitted, and model parameters are effectively compressed by means of quantization, sparsification, etc., reducing communication bandwidth requirements. On the other hand, model distillation technology is used to extract key information and simplify complex global models into lightweight local models while maintaining the recommended performance of the model, thereby further reducing communication costs. In addition, the present invention also designs efficient model parameter update strategies, such as asynchronous updates or gradient-based updates, to reduce communication rounds and waiting time, and improve the overall efficiency of federated learning.

[0038] On the basis of model parameter encryption and transmission, the present invention further introduces a higher level of privacy protection mechanism. First, differential privacy technology is used to add noise to the model parameters to ensure that even if the model parameters are leaked during transmission, the specific user data cannot be inferred, thereby effectively protecting user privacy. Secondly, the distributed storage and verification mechanism based on blockchain is explored, and the decentralization and immutability of blockchain are used to enhance the transparency and security of model parameter transmission and aggregation. These measures together constitute a strict privacy protection system, which provides a strong guarantee for the safe and stable operation of the federated recommendation system.

[0039] In order to improve the robustness of the federated recommendation system, the present invention introduces adversarial training or robustness optimization technology. By introducing adversarial samples or noise, the training model can still maintain good performance when facing data disturbances or malicious attacks. This training method can enhance the generalization ability and anti-attack ability of the model, making it more stable and reliable in the face of complex and changeable network environments. In addition, the present invention also designs anomaly detection and fault-tolerant mechanisms, which can promptly detect and handle potential problems such as client failures or data pollution, and ensure the continuous and stable operation of the system. These measures jointly improve the overall robustness and reliability of the federated recommendation system.

[0040] See also Figure 1-Figure 4 ,In the embodiment of the present invention, in step (i), on the basis of the global model, each client can make fine adjustments according to the uniqueness of local data; this includes adjusting the architecture of the model, optimizing the objective function, or introducing additional regularization items to adapt to different user groups or application scenarios; In step (ii), before the model parameters are transmitted, the model parameters are preprocessed using a parameter compression algorithm to reduce the amount of data transmitted, which includes methods such as quantizing parameter values, sparse representation, or using low-rank decomposition; by designing efficient communication protocols and scheduling strategies to minimize communication delays and bandwidth usage; In step (iii), on the basis of encrypting the model parameters, differential privacy technology is further used to protect user privacy; by adding an appropriate amount of noise to the model parameters, it is ensured that the data of a single user cannot be accurately inferred during the model training process; blockchain technology is introduced to achieve distributed storage and verification of model parameters; the security and credibility of model parameters are ensured through the immutability and decentralization characteristics of blockchain; In step (iv), adversarial training techniques are used to train the model by introducing adversarial samples or noise to improve the model's robustness to input data perturbations; anomaly detection and fault-tolerance mechanisms are designed to detect and respond to potential problems such as client failures, data contamination, or malicious attacks.

[0041] The working principle of the embodiment of the present invention is: after the global model is distributed to each client, the present invention allows each client to fine-tune the global model according to the uniqueness of the local data. This fine-tuning is not limited to adjusting the hyperparameters of the model, but also includes optimizing the model architecture, adjusting the objective function, or introducing additional regularization terms to adapt to different user groups or application scenarios. This personalized adjustment strategy can make full use of the local data advantages of each client, making the model more in line with the preferences of a specific user group, thereby improving the recommendation effect and user satisfaction.

[0042] During the model parameter transmission process, the present invention uses a parameter compression algorithm to preprocess the model parameters to reduce the amount of data transmitted. These compression algorithms include quantizing parameter values, sparse representation, or using low-rank decomposition, which can significantly reduce the communication bandwidth requirements. At the same time, by designing efficient communication protocols and scheduling strategies, the present invention further minimizes communication delays and bandwidth occupancy, and improves the overall efficiency of federated learning.

[0043] Working principle: First, user behavior data is collected through multiple distributed clients, and then cleaned and feature extracted to construct a user-item interaction graph to provide effective data input for graph neural network model training. Then, the graph neural network model is used for local training on each client to learn the potential representation of users and items, and optimize the model parameters to predict user preferences. This step enables each client to train a preliminary recommendation model based on local data. Then, the trained model parameters are encrypted to ensure privacy and security during data transmission. The encrypted model parameters are transmitted to the central server for aggregation to form a global model. During the aggregation process, the present invention adopts a federal average algorithm and considers the data volume and model training quality of each client to ensure the accuracy and robustness of the global model. After the global model is distributed to each client, each client can fine-tune or customize the global model according to the uniqueness of the local data to further improve the recommendation effect and user satisfaction. This personalized adjustment strategy makes full use of the local data advantages of each client. In addition, the present invention also considers aspects such as communication efficiency optimization, privacy protection enhancement, and system robustness improvement. By adopting parameter compression algorithms, differential privacy technology and blockchain technology, communication overhead is reduced, privacy protection is enhanced, and the security and credibility of the system are improved. At the same time, the introduction of adversarial training technology and anomaly detection mechanism improves the robustness of the model to input data disturbances, ensuring the continuous and stable operation of the system.

[0044] In summary, the working principle of the embodiment of the present invention realizes efficient model training and optimization using data from multiple distributed clients while protecting user privacy, providing a new solution for the recommendation system.

[0045] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit thereof, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A research method for the aggregation strategy of the federated recommendation system model based on graph neural network, characterized in that: include: Step (1), data preprocessing stage, is used to collect and clean data and extract features from the cleaned data; Step (2), local model training phase, using the graph neural network model for local training, optimizing model parameters to predict user preferences; Step (3), model parameter encryption and transmission stage, encrypt the GNN model parameters trained on each client; Step (4), the federated aggregation phase, receives the encrypted model parameters from each client, performs decryption operations, and applies federated averaging or other advanced federated learning aggregation algorithms; Step (5), the global model update and distribution phase, re-encrypts the aggregated global model parameters and distributes them back to each client; Step (6), iterative optimization and evaluation phase, performs iterative optimization until the preset training rounds or the model performance improvement is no longer significant.

2. According to claim 1, a research method for aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: The step (1) comprises the following steps: a. Collect user behavior data on multiple distributed clients, including but not limited to user click, purchase, evaluation and other historical records; b. Clean the collected data, remove invalid, duplicate or abnormal data, and ensure data quality; c. Perform feature extraction on the cleaned data and construct a user-item interaction graph, where nodes represent users or items and edges represent the interaction between users and items.

3. According to claim 1, a research method for aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: The step (2) comprises the following steps: a. On each client, based on the constructed user-item interaction graph, a graph neural network (GNN) model is used for local training to learn the potential representations of users and items; b. During local training, a graph convolutional network (GCN), graph attention network (GAT), or other graph neural network variants are used to combine user historical behavior and item attributes to optimize model parameters to predict user preferences.

4. According to claim 1, a research method for aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: The step (3) comprises the following steps: a. Encrypt the GNN model parameters trained on each client to ensure data privacy and security during transmission; b. Transmit the encrypted model parameters to the central server for subsequent model aggregation; The encryption process in step (3) adopts homomorphic encryption, differential privacy or secure multi-party computing technology to ensure the security of data transmission.

5. According to claim 1, a method for researching aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: The step (4) comprises the following steps: a. The central server receives the encryption model parameters from each client and performs decryption operations; b. Apply Federated Averaging or other advanced federated learning aggregation algorithms to aggregate the decrypted model parameters to form a global model; c. Optionally, introduce model regularization or weight decay techniques to improve the generalization ability of the global model; The federated averaging algorithm in step (4) further considers the amount of data and model training quality of each client, and adopts a weighted averaging strategy for model aggregation.

6. According to claim 1, a research method for aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: The step (5) comprises the following steps: a. Re-encrypt the aggregated global model parameters and distribute them back to each client; b. Each client receives the updated global model parameters, replaces the local model, and prepares for the next round of training; In step (5), when the global model is updated, the stability and convergence speed of the model are also considered, and the learning rate is dynamically adjusted or other optimization strategies are adopted.

7. According to claim 1, a method for researching aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: The step (6) comprises the following steps: a. Repeat steps (2) to (5) until the preset number of training rounds is reached or the model performance is no longer significantly improved; b. After each round of iteration, the performance of the global model is evaluated, including common indicators of recommendation systems such as accuracy, recall, and F1 score.

8. A method for studying aggregation strategies of federated recommendation system models based on graph neural networks according to any one of claims 1 to 7, characterized in that: The following optional steps are also included: Step (i). Model personalization adjustment stage: After the global model is distributed to each client, the global model is fine-tuned or customized according to the specific needs of each client, user group characteristics or local data set characteristics to further improve the recommendation effect and user satisfaction; this step involves adjusting the model's learning rate, regularization parameters, or introducing additional local features for training; Step (ii). Communication efficiency optimization stage. Considering the communication overhead in federated learning, this method can further optimize the transmission process of model parameters; use parameter compression technology to reduce the amount of transmitted data, or use model distillation technology to extract key information to reduce communication costs; Design efficient model parameter update strategies, either asynchronously or based on gradients, to reduce communication rounds and waiting time; Step (iii). In the privacy protection enhancement stage, based on the encryption and transmission of model parameters, a higher level of privacy protection mechanism is further introduced; by using differential privacy technology to add noise to the model parameters, it is ensured that even if the model parameters are leaked, the specific user data cannot be inferred; at the same time, a distributed storage and verification mechanism based on blockchain can be explored to enhance the transparency and security of the model parameter transmission and aggregation process; Step (iv). System robustness improvement stage: In order to improve the robustness of the federated recommendation system, this method can introduce adversarial training or robustness optimization technology; By introducing adversarial samples or noise, the training model can still maintain good performance when facing data perturbations or malicious attacks; in addition, anomaly detection and fault-tolerant mechanisms are designed to deal with potential problems such as client failures or data pollution.

9. According to claim 8, a method for researching aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: In step (i), based on the global model, each client can perform fine-tuning according to the uniqueness of local data; this includes adjusting the model architecture, optimizing the objective function, or introducing additional regularization terms to adapt to different user groups or application scenarios; In the step (ii), before the model parameters are transmitted, a parameter compression algorithm is used to preprocess the model parameters to reduce the amount of data transmitted, which includes quantizing parameter values, sparse representation or using low-rank decomposition and the like; and by designing efficient communication protocols and scheduling strategies to minimize communication delays and bandwidth occupancy.

10. According to claim 8, a method for researching aggregation strategy of federated recommendation system model based on graph neural network is characterized in that: In the step (iii), on the basis of encrypting the model parameters, differential privacy technology is further used to protect user privacy; by adding an appropriate amount of noise to the model parameters, it is ensured that the data of a single user cannot be accurately inferred during the model training process; blockchain technology is introduced to achieve distributed storage and verification of model parameters; and the security and credibility of model parameters are ensured through the immutability and decentralization characteristics of blockchain; In the step (iv), adversarial training technology is used to train the model by introducing adversarial samples or noise to improve the robustness of the model to input data disturbances; anomaly detection and fault tolerance mechanisms are designed to detect and respond to potential problems such as client failures, data contamination or malicious attacks.

Images