6G edge network user prediction method based on fairness federation

Abstract

The invention discloses a 6G edge network user prediction method based on fairness federation. The method comprises the following steps: step 1, initializing a 6G edge network user prediction mechanism based on fairness federation; 2, the edge core network completes local model updating according to a personalized enhancement strategy; 3, the central core network integrates knowledge to construct a global model according to a fairness aggregation strategy; 4, based on local model updating and a global model, training a 6G edge network user prediction mechanism based on fairness federation; and 5, performing 6G user prediction by using the trained 6G edge network user prediction mechanism based on the fairness federation. The method has the characteristics of personalized service enhancement, fair strategy implementation, prediction performance improvement and data privacy protection.

Description

A Fair Federated Approach for 6G Edge Network User Prediction Technical Field

[0001] This invention belongs to the field of 6G user prediction methodology, specifically relating to a 6G edge network user prediction method based on fair federation. Background Technology

[0002] The 6th Generation Mobile Networks (6G) system, in its drive for "integrated communication, sensing, and computing," deeply integrates technologies such as intelligent metasurfaces and artificial intelligence, and connects to diverse terminals including satellites, automobiles, and wearable devices. With the explosive growth in the number of terminals and the volume of sensor data, the network edge faces unprecedented processing pressure. This pressure is driving the transformation of 6G edge networks from passive response to proactive prediction. Faced with the contradiction between dynamically changing user needs and limited edge resources, predicting user behavior has become crucial for breaking the deadlock. In Release 18, the Third Generation Partnership Project (3GPP) addressed the issue of missing GNSS signals in indoor environments by using artificial intelligence (AI) algorithms to identify the line-of-sight (LOS) propagation path between base stations (gNodeBs, gNBs) and user equipment (UEs). Based on this achievement, 3GPP plans to build an AI-enabled intelligent positioning technology framework in Release 19, advocating for user prediction. However, traditional centralized architectures are difficult to adapt to the scalability requirements of 6G ultra-large-scale networking scenarios.

[0003] Federated Learning (FL), as a distributed machine learning paradigm, offers a feasible solution to the aforementioned problems. FL allows edge core networks to complete model training at the edge and share only model parameters without transmitting private data to the central core network. This characteristic aligns well with the stringent user privacy protection requirements of 6G networks and enhances network scalability. However, the regional specificity of end-user data in 6G edge environments leads to heterogeneous characteristics such as non-independent and identically distributed (Non-IID) data and uneven data volume. This heterogeneity presents two major challenges for Federated Learning in 6G edge networks. First, traditional global models struggle to effectively accommodate the personalized quality of service requirements of all edge core networks. Second, weighted aggregation mechanisms based on sample size tend to suppress the contribution of models with smaller data domains, resulting in decreased aggregation fairness and uneven global model performance. Therefore, optimizing the training and aggregation strategies of Federated Learning to adapt to the high heterogeneity of 6G edge networks is crucial for achieving its performance goals.

[0004] In 6G scenarios, scholars are exploring the application of federated learning. Zhou et al. proposed a two-layer federated learning algorithm for vehicle detection in 6G vehicle-to-everything (V2X) environments, using vehicle position, speed, and direction of motion as indicators to construct aggregate weights, effectively improving training efficiency. Chen et al. considered combining federated learning with a Low Earth Orbit Satellite (LEO) system, demonstrating that the proposed federated method possesses practical accuracy. Zhou et al. designed a federated learning algorithm based on decomposition and meta-deep reinforcement learning, aiming to adapt to dynamic satellite-ground environments to achieve efficient upload and aggregation processes.

[0005] A patent application (CN119854824A) discloses an efficient training optimization method for federated learning in 6G networks, proposing a federated learning framework to address the communication efficiency problem in 6G networks. This method adds constraint terms to the client's objective function to improve convergence speed and reduce communication rounds. Simultaneously, it designs rules for partial client participation, prioritizing computationally efficient clients for uploading parameters, thereby reducing the cost per round of communication and improving the training efficiency and model performance of federated learning.

[0006] A patent with publication number CN120224300A discloses a federated learning method, apparatus, and device for integrated air-space-ground networks, proposing a multi-layered federated learning architecture and collaborative training mechanism suitable for integrated air-space-ground networks. This method constructs a three-tiered computing and communication system of "satellite-UAV-ground user," utilizing UAVs as edge servers to aggregate user-local models and leveraging near-Earth orbit satellites for global aggregation in the cloud.

[0007] Federated learning has made some progress in 6G research, demonstrating its performance benefits in 6G scenarios. However, these explorations generally suffer from two limitations. Firstly, the validation phase relies heavily on general datasets like MNIST, lacking support from real-world mobile communication network data, and the introduction of domain-specific metrics limits its scalability to other scenarios. Secondly, due to the significant non-independent and identically distributed nature and regional heterogeneity of user data, traditional federated aggregation methods face deeper challenges. Specifically, traditional federated learning, aiming to build a unified global model, often ignores the differences in behavioral patterns between regions, making it difficult to provide effective personalized services. More seriously, the uneven data distribution among terminals inherently flaws aggregation mechanisms based on data volume weighting, leading to devices with large data volumes dominating, triggering a "Matthew effect," resulting in unfair aggregation, unstable performance, and weakened model generalization ability and prediction fairness. Furthermore, the lack of in-depth design of collaborative mechanisms between the global and local models results in low knowledge integration efficiency in existing methods, limiting prediction accuracy. Summary of the Invention

[0008] To overcome the shortcomings of the existing technology, the present invention aims to provide a 6G edge network user prediction method based on fair federation, which features enhanced personalized services, implementation of fair policies, improved prediction performance, and protection of data privacy.

[0009] To achieve the above objectives, the technical solution adopted by this invention is: a 6G edge network user prediction method based on fair federation, comprising the following steps: Step 1: Initialize the 6G edge network user prediction mechanism based on fair federation. Step 2: The edge core network is enhanced according to a personalized strategy. Step 3: The central core network aggregates according to the fairness strategy, completing the local model update; Step 4: Based on the local model update and the global model, train a 6G edge network user prediction mechanism based on fairness federation. Step 5: Utilize the trained 6G edge network user prediction mechanism based on fairness federation. 6G user forecasting.

[0010] First, execute step 1 to complete the initialization. Then, steps 2 (local model update) and 3 (global model aggregation) are executed iteratively until the maximum number of global communication rounds is reached, thus completing the training mechanism shown in step 4. Finally, the trained communication models and local personalized models of each edge core network are deployed in the edge core network, and step 5 is executed for user prediction.

[0011] Specifically, the prediction results are sent to the Access and Mobility Management Function (AMF) via the Policy Control Function (PCF). The AMF then interacts with the gNB that is expected to be connected through the N2 interface to achieve accurate analysis and real-time prediction of user data.

[0012] Step 1 specifically includes: 1.1): First, preprocessing the UE movement trajectory information collected by the AMF according to the time window. The historical trajectories of the UE are divided into different gNBs; whereby user input data is represented by a time index. The actual prediction results are represented by the time index as follows: ,and A single-step prediction strategy is adopted, that is, based on historical data. Step trajectory prediction of the target gNB location where the user will switch at the next moment; 1.2): Introducing the Dirichlet Distribution (DD) to simulate the Non-IID characteristics of user data in massive 6G edge networks, expressed as the following formula: ,in, It is a probability vector that satisfies and , Used to characterize the degree of heterogeneity of data The smaller the value, the higher the heterogeneity of user data within each edge core network. It is a multivariate beta function; 1.3): Initialize the 6G user prediction model deployed in each core network, and select the Long Short-Term Memory (LSTM) model as the basic model for implementing 6G user prediction; the basic model is deployed in the Network Data Analytics Function (NWDAF) of the central core network and the edge core network respectively; at the same time, the global user prediction model parameters initialized in the central core network are... and issued to Each edge core network completes its local model. to The initialization work is performed; during model training, mean square error (MSE) is used to measure the difference between the prediction and the actual situation. The sample-level loss is shown below: ,in, and These represent the actual base stations and the predicted results, respectively. For edge core network User prediction model, It is an edge core network The parameters of the user prediction model; to record the performance of the designed model under a specific data distribution, a client-level loss function is designed based on the data used for training the current model: ;1.4): Initialize the framework optimization objective; Based on the local loss function defined in 1.3), the overall optimization objective of the federated learning framework is also established. This objective aims to unify the contributions of all edge and core networks in a weighted manner, while reasonably reflecting the differences in the contribution of their edge and core network models, and is defined as follows: ,in, The aggregate weights for each edge core network, , The model contribution weights for each edge core network.

[0013] In step 1.2), the present invention adjusts based on Dirichlet distribution. Take values, simulate and generate Highly heterogeneous dataset distribution in edge core networks This framework is designed for a single central core network and Each edge core network; Having a private dataset is represented as Consistent with distribution , ,and For dataset The number of samples in the core network; among them, the edge core network users The input trajectory data is represented as The actual predicted base station data is represented as .

[0014] In section 1.3), when calculating accuracy, the prediction results of the selected user prediction model and the actual gNB coordinates are placed on a KD tree, and the nearest gNB index information is returned; if the two indices are the same, the prediction is considered correct. Conversely, it means the prediction was wrong. At the same time, the mean is used to calculate the accuracy.

[0015] Step 2 specifically includes: 2.1): Constructing a personalized enhancement strategy. For edge core network 2.2): Constructing communication models and local personalized models for the edge core network. The Model Training Logical Function (MTLF) constructs the communication model used for aggregation. , communication model The loss function is shown below: ,in, The loss is the MSE loss obtained from training the 6G user prediction model. Composition; 2.3): edge core network MTLF builds a personalized model that is persistently retained locally. Local personalized model The loss function is shown below: , ,in, The loss is the MSE loss obtained from model training. and the penalty term loss used to close the distance composition, Indicates the penalty coefficient; 2.4): Edge core network The communication model and the local personalization model are trained separately in MTLF, as shown below: , ,in, and They represent the first Wheel and First The core network of the wheel's edge Communication model parameters, and They represent the first Wheel and First The core network of the wheel's edge Local personalized model parameters, and These represent the edge and core networks, respectively. The learning rates of the communication model and the local personalized model; simultaneously, training stops after reaching the maximum number of training rounds, and the communication model... The data is uploaded to the central core network; furthermore, the prediction results of the two models are expressed as follows: and .

[0016] Step 3 specifically includes: 3.1): Constructing a fair aggregation strategy. 3.2): User data generated through Dirichlet distribution simulation exhibits high heterogeneity, with significant differences in data distribution between different edge core networks. Therefore, this invention designs a fair aggregation strategy. The aim is to introduce the Fisher Information Matrix (FIM) to improve the fairness of global model knowledge coverage by aggregating weights through multiple indicators. FIM is an important concept in information theory, usually used to quantify the identifiability and information content of model parameters. Its calculation formula is shown below: The cumulative sum of FIM parameters is selected as the importance index for the model parameters and is calculated by the following formula: ,in, For communication model The total number of parameters; 3.3): Communication model of each edge core network in each round. After the update is complete, utilize its parameter importance indicators. and sample size Calculate the weights of multidimensional indicators And uploaded to the central core network: ,in, For hyperparameters, the larger the value... The value represents the importance of the model parameters. The polymerization effect is more obvious; when When, the strategy is equal to FedAvg; when At that time, the weights were determined entirely by the importance of the model parameters. To allocate; 3.4): The server utilizes multi-dimensional aggregation parameters The aggregation process is completed according to the following formula: , ,in, The aggregate weights for each edge core network, To update Global model at time step Indicates the number of communication rounds. Edge Core Network The communication model parameters; 3.5): The global model formed by the aggregation of the central core network. It is distributed to each edge core network.

[0017] Step 4 specifically involves: using a 6G edge network user prediction mechanism based on fair federation. In this process, the training process iterates continuously between steps 2 and 3 until the maximum number of global communication rounds is reached. The Adam optimizer is used to train the communication models and personalized models of each edge core network.

[0018] Step 5 specifically involves: the user trajectory data undergoing the 6G edge network user prediction mechanism based on fairness federation from step 4. After training, the communication models and local personalized models of each edge core network are deployed in its analytical logic function (AnLF). The prediction results are pushed to the PCF through the Nnwdaf_AnalyticsInfo service. After receiving the prediction, the PCF issues mobility optimization strategies to the AMF. Subsequently, the AMF pre-interacts with the gNB to be connected through the N2 interface to achieve accurate analysis and real-time prediction of user data.

[0019] The beneficial effects of this invention are as follows: First, this invention constructs a personalized enhancement strategy. This strategy considers the Non-IID characteristics of user data, optimizes locally, enhances the correlation between the global model and the local model, promotes more effective learning and integration of knowledge information by the personalized model, and helps to accelerate model convergence.

[0020] Second, this invention constructs a fair aggregation strategy. This strategy introduces the Fisher Information Matrix (FIM) into the aggregation part, and with the help of indicators such as data volume, it provides multi-dimensional metrics for aggregation, thereby achieving more globally representative model aggregation and enabling the model to exhibit better predictive performance. Attached Figure Description

[0021] Figure 1 is a flowchart of the implementation of the present invention.

[0022] Figure 2 is a schematic diagram of the 6G edge network user prediction mechanism based on fair federation of the present invention.

[0023] Figure 3 is a comparison of the effects of the present invention and the traditional federated algorithm. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings.

[0025] As shown in Figure 1, a 6G edge network user prediction method based on fair federation includes the following steps: Step 1 specifically includes: 1.1): First, the UE movement trajectory information collected by the AMF is preprocessed according to the time window. Divide the historical trajectory of the UE in different gNBs.

[0026] The user input data is represented by a time index. The actual prediction results are represented by the time index as follows: ,and This invention employs a single-step prediction strategy, that is, based on historical data. The gait trajectory predicts the target gNB position that the user will switch to next.

[0027] 1.2): To simulate the Non-IID characteristics of user data in massive 6G edge networks, the Dirichlet Distribution (DD) is introduced. The Dirichlet Distribution is a continuous multivariate probability distribution widely used in parameter estimation and Bayesian inference for probabilistic models, and can be expressed as follows: ,in, It is a probability vector that satisfies and , Used to characterize the degree of heterogeneity of data It is a multivariate beta function.

[0028] In this invention, the following are selected , The smaller the value, the higher the heterogeneity of the data within each edge network. This invention is based on Dirichlet distribution adjustment. Take values, simulate and generate Highly heterogeneous dataset distribution in edge core networks This framework is designed for a single central core network and Each edge core network. Having a private dataset is represented as Consistent with distribution , ,and For dataset The number of samples in the core network. Among them, the edge core network... users The input trajectory data is represented as The actual predicted base station data is represented as .

[0029] 1.3): Initialize the 6G user prediction model deployed in each core network. A Long Short-Term Memory (LSTM) model is selected as the base model for 6G user prediction. This model is deployed in the NWDAF of both the central core network and the edge core networks. Simultaneously, the global user prediction model parameters are initialized in the central core network. Distribute to Each edge core network completes its local model. to Initialization work.

[0030] When calculating accuracy, the prediction results of the selected user prediction model and the actual gNB coordinates are plotted on a KD-tree, and the nearest neighbor gNB index is returned. If the two indices are the same, the prediction is considered correct. Conversely, it means the prediction was wrong. At the same time, the mean is used to calculate the accuracy.

[0031] During model training, we use Mean Square Error (MSE) to measure the difference between the prediction and the actual situation. The sample-level loss is shown below: ,in, and These represent the actual base stations and the predicted results, respectively. For edge core network User prediction model, It is an edge core network The parameters of the user prediction model.

[0032] To record the performance of the designed model under a specific data distribution, a client-level loss function is designed based on the data used for training the current model: 1.4): Initializing the Framework Optimization Objective; Based on the local loss function defined in 1.3), the overall optimization objective of the federated learning framework is also established. The optimization objective of this invention is to train a global model that shares all knowledge through distributed collaborative training while protecting user data from leaving their homes. To achieve this objective, the overall optimization function is designed as a weighted form to reflect the differences in data volume among different participants and their contributions to the global model, defined as follows: ,in, As the aggregate weight of each edge core network, in this invention, , The model contribution weights for each core network.

[0033] Step 2 specifically involves: 2.1) In federated learning, the sample-weighted method performs well with IID data, but its applicability is significantly reduced in the heterogeneous data environment of 6G edge networks. To address this issue, as shown in Figure 2, this invention proposes a personalized enhancement strategy. This involves building communication models and local personalized models for the edge core network, enhancing the correlation between the global model and the local model, improving model performance, and increasing the ability to provide high-quality services to users.

[0034] 2.2): For the edge core network The MTLF is used to construct a communication model for participating in aggregation. This model serves as the medium for interaction between the core network and the central core network, and is periodically reset to a global model. It is based on MTLF, which uses the Nadrf interface to obtain historical user data from the Analytics Data Repository Function (ADRF) for model training, and then uploads it to the central core network for global aggregation. (Communication Model) The loss function is as follows: ,in, The loss is the MSE loss obtained from training the 6G user prediction model. composition.

[0035] 2.3): For the edge core network MTLF builds a personalized model that is persistently retained locally. Personalized models A penalty term is introduced to narrow the distance between the local personalized model and the global model, achieving a dynamic balance between local knowledge retention and global knowledge penetration. This ensures... Without incurring additional communication overhead, this model maximizes the retention of local knowledge while selectively absorbing global knowledge, achieving superior prediction results. Local Personalized Model The loss function is as follows: , ,in, The loss is the MSE loss obtained from model training. and the penalty term loss used to close the distance composition, This represents the penalty coefficient.

[0036] 2.4): Edge Core Network The communication model and the local personalization model are trained separately in MTLF, as shown below: , ,in, and They represent the first Wheel and First The core network of the wheel's edge Communication model parameters, and They represent the first Wheel and First The core network of the wheel's edge Local personalized model parameters, and These represent the edge and core networks, respectively. The learning rates of the communication model and the local personalized model. Simultaneously, training stops after reaching the maximum number of training epochs, and the communication model... It is uploaded to the central core network. Furthermore, the prediction results of the two models are expressed as follows: and .

[0037] Step 3 specifically involves: 3.1) When data distributions differ significantly, data aggregation methods can weaken the contributions of smaller participants, hindering the global model from fully absorbing the knowledge of all participants. Therefore, this invention designs a fair aggregation strategy. As shown in Figure 2, the Fisher information matrix is ​​introduced, and the weights are aggregated through multiple indicators to improve the fairness of the global model knowledge coverage.

[0038] 3.2): FIM is an important concept in information theory, usually used to quantify the identifiability and information content of model parameters. Its calculation formula is shown below: The sum of FIM parameters is selected as the importance index for the model parameters and is calculated by the following formula: ,in, For communication model The total number of parameters.

[0039] 3.3): Communication model of each edge core network in each round After the update is complete, utilize its parameter importance indicators. and sample size Calculate the weights of multidimensional indicators And uploaded to the central core network: ,in, For hyperparameters, the larger the value... The value represents the importance of the model parameters. The polymerization effect is more pronounced. When When, this strategy is equal to FedAvg. At that time, the weights were determined entirely by the importance of the model parameters. To allocate.

[0040] 3.4): The server utilizes multi-dimensional aggregation parameters The aggregation process is completed according to the following formula: , ,in, The aggregate weights for each edge core network, To update Global model at time step Indicates the number of communication rounds. Edge Core Network The communication model parameters.

[0041] 3.5): The central core network will aggregate to form a global model. It is distributed to each edge core network.

[0042] Step 4 specifically involves: using a 6G edge network user prediction mechanism based on fair federation. In this process, the training process iterates continuously between steps 2 and 3 until the maximum number of global communication rounds is reached, and the Adam optimizer is used to train two models for each edge core network.

[0043] Step 5 specifically involves: the user trajectory data being processed by the 6G edge network user prediction mechanism based on fairness federation, as described in step 4. After training, the communication models and local personalized models of each edge core network are deployed in its AnLF. The prediction results are pushed to the PCF through the Nnwdaf_AnalyticsInfo service. After receiving the prediction, the PCF issues mobility optimization strategies to the AMF. The AMF pre-interacts with the gNB to be connected through the N2 interface to achieve accurate analysis and real-time prediction of user data.

[0044] I. Experimental Conditions The simulation experiment of this invention was conducted on Windows 10 Professional Chinese Edition, configured with a 4.30GHz Intel Core™ i5-10400 CPU, an NVIDIA GeForce RTX 4090 GPU with 24GB of video memory, the deep learning framework used was PyTorch 2.2.1, and the programming language was Python 3.9.

[0045] The dataset used in the simulation experiment of this invention is a real user movement trajectory dataset provided by Shanghai Telecom. Each record includes the user's ID, start time, end time, and gNB latitude and longitude coordinates. This dataset contains more than 7.2 million records, including records of 9,481 terminal UEs accessing the network through 3,233 gNBs within 6 months.

[0046] II. Experimental Content and Analysis To verify the effectiveness of this invention, accuracy was selected as the evaluation index to assess the predictive performance of the model. Simultaneously, to demonstrate the designed 6G edge network user prediction mechanism based on fair federation... To assess its advancements, we compare it with four commonly used federated algorithms.

[0047] The four commonly used federated algorithms include: Pure local non-federated: This means that each edge core network uses only the local dataset for model training and does not participate in the federated learning process.

[0048] FedAvg is a pioneering work in federated learning, which uses the amount of data to guide the aggregation of local models.

[0049] FedProx: Uses regularization terms to constrain updates from edge devices; hyperparameter set to 1.

[0050] Ditto: Uses a regularization term to maintain the correlation between the local model and the global model; the hyperparameter is set to 1.

[0051] Figure 3 shows the accuracy comparison of the present invention with four baseline methods on the test set. To verify the effectiveness of the method, the experiment used a Dirichlet distribution to simulate heterogeneous data distribution, setting α to 0.1, 0.3, 0.5, 0.7, and 1 respectively to complete the partitioning of different heterogeneous datasets. It is worth noting that the present invention achieved optimal performance under the current experimental configuration.

[0052] Specifically, when α = 1, the accuracy of this invention is 64.94%. When α = 0.1, the accuracy is 61.58%, which is 4.63%, 2.71%, 1.89%, and 1.41% higher than the pure local non-federated scheme, FedAvg, FedProx, and Ditto, respectively. This verifies the effectiveness of the personalized enhancement strategy and the fairness aggregation strategy proposed in this paper. The increase in prediction accuracy indicates that the designed local personalized model can achieve a balance between local personalized knowledge and global knowledge, providing users with customized and high-quality personalized services. At the same time, the designed fairness aggregation strategy can adaptively adjust the aggregation weights. When the model parameters of the current edge core network are more important, the system will increase their influence in aggregation. When their importance is relatively reduced, the framework will rely more on the amount of data and other client importance indicators for comprehensive judgment. This multi-indicator collaborative weight allocation mechanism can more effectively capture global knowledge and improve the model's learning effect.

[0053] The purpose of this invention is to provide a 6G edge network user prediction mechanism based on fair federation. First, a federated learning architecture avoids the transmission of raw data across devices, strengthening privacy protection at the architectural level and improving support for ultra-large-scale terminal access. Simultaneously, a personalization enhancement strategy is introduced to improve the local model's ability to represent Non-IID data, thereby enhancing its adaptability to region-specific patterns and the level of personalized service. Furthermore, a fairness aggregation mechanism based on multi-dimensional evaluation metrics such as the Fisher information matrix is ​​designed to mitigate model bias caused by data imbalance, comprehensively improving the generalization and fairness of the aggregated model. Finally, by strengthening the collaboration and knowledge integration between the global model and the local model, prediction accuracy and user experience are significantly improved while supporting personalized services.

Claims

1. A 6G edge network user prediction method based on fair federation, characterized in that, Includes the following steps; Step 1: Initialize a 6G edge network user prediction mechanism based on fairness federation Step 2: The edge core network is enhanced according to a personalized strategy. Step 3: The central core network aggregates according to the fairness strategy, completing the local model update; Step 4: Based on the local model update and the global model, train a 6G edge network user prediction mechanism based on fairness federation. Step 5: Utilize the trained 6G edge network user prediction mechanism based on fairness federation. 6G user forecasting.

2. The 6G edge network user prediction method based on fair federation according to claim 1, characterized in that, Step 1 specifically includes: 1.1): First, preprocessing the UE movement trajectory information collected by the AMF according to the time window. The historical trajectories of the UE are divided into different gNBs; whereby user input data is represented by a time index. The actual prediction results are represented by the time index as follows: ,and A single-step prediction strategy is adopted, that is, based on historical data. The gait trajectory predicts the target gNB location where the user will switch at the next moment; 1.2): Introducing the Dirichlet distribution (DD) to simulate the Non-IID characteristics of user data in massive 6G edge networks, expressed as the following formula: ,in, It is a probability vector that satisfies and ， Used to characterize the degree of heterogeneity of data The smaller the value, the higher the heterogeneity of user data within each edge core network. It is a multivariate beta function; 1.3): Initialize the 6G user prediction model deployed in each core network, and select the Long Short-Term Memory (LSTM) model as the basic model for implementing 6G user prediction; the basic model is deployed in the network data analysis function network elements of the central core network and the edge core network respectively; at the same time, the global user prediction model parameters initialized in the central core network are... and issued to Each edge core network completes its local model. to The initialization work is performed; during model training, mean squared error is used to measure the difference between the prediction and the actual situation, and the sample-level loss is shown below: ,in, and These represent the actual base stations and the predicted results, respectively. For edge core network User prediction model, It is an edge core network The parameters of the user prediction model; based on the data used for training the current model, design a client-level loss function and record the performance of the designed model under a specific data distribution: ;1.4): Initialize the framework optimization objective; Based on the local loss function defined in 1.3), the overall optimization objective of the federated learning framework is also established. This objective aims to unify the contributions of all edge and core networks in a weighted manner, while reasonably reflecting the differences in the contribution of their edge and core network models, and is defined as follows: ,in, The aggregate weights for each edge core network, ， The model contribution weights for each edge core network.

3. The 6G edge network user prediction method based on fair federation according to claim 2, characterized in that, In step 1.2), the adjustment is based on the Dirichlet distribution. Take values, simulate and generate Highly heterogeneous dataset distribution in edge core networks This framework is designed for a single central core network and Each edge core network; Having a private dataset is represented as Consistent with distribution ， ,and For dataset The number of samples in the core network; among them, the edge core network users The input trajectory data is represented as The actual predicted base station data is represented as 。 4. The 6G edge network user prediction method based on fair federation according to claim 2, characterized in that, In section 1.3), when calculating the accuracy, the prediction results of the selected user prediction model and the actual gNB coordinates are placed on the KD tree, and the gNB index information of the nearest neighbor between the two is returned. If the two indexes are the same, the prediction is considered correct. Conversely, it means the prediction was wrong. At the same time, the mean is used to calculate the accuracy.

5. The 6G edge network user prediction method based on fair federation according to claim 2, characterized in that, Step 2 specifically includes: 2.1): Constructing a personalized enhancement strategy. For edge core network 2.2): Constructing communication models and local personalized models for the edge core network. The model training logic function constructs a communication model for participating in aggregation. , communication model The loss function is shown below: ;in, The loss is the MSE loss obtained from training the 6G user prediction model. Composition; 2.3): edge core network MTLF builds a personalized model that is persistently retained locally. Local personalized model The loss function is shown below: ， ,in, The loss is the MSE loss obtained from model training. and the penalty loss used to close the distance composition, Indicates the penalty coefficient; 2.4): Edge core network The communication model and the local personalization model are trained separately in MTLF, as shown below: ， ,in, and They represent the first Wheel and First The core network of the wheel's edge Communication model parameters, and They represent the first Wheel and First The core network of the wheel's edge Local personalized model parameters, and These represent the edge and core networks, respectively. The learning rates of the communication model and the local personalized model; simultaneously, training stops after reaching the maximum number of training rounds, and the communication model... The data is uploaded to the central core network; furthermore, the prediction results of the two models are expressed as follows: and 。 6. The 6G edge network user prediction method based on fair federation according to claim 5, characterized in that, Step 3 specifically includes: 3.1): Constructing a fair aggregation strategy. 3.2): Integrating knowledge in the central core network to construct a global model; The user data generated through Dirichlet distribution simulation exhibits high heterogeneity, so a fair aggregation strategy should be designed. Fisher's Information Matrix (FIM) is introduced to improve the fairness of global model knowledge coverage by aggregating weights through multiple indicators; its calculation formula is shown below: The cumulative sum of FIM parameters is selected as the importance index for the model parameters and is calculated by the following formula: ,in, For communication model The total number of parameters; 3.3): Communication model of each edge core network in each round. After the update is complete, utilize its parameter importance indicators. and sample size Calculate the weights of multidimensional indicators And uploaded to the central core network: ,in, For hyperparameters, the larger the value... The value represents the importance of the model parameters. The polymerization effect is more obvious; when When, the strategy is equal to FedAvg; when At that time, the weights were determined entirely by the importance of the model parameters. To allocate; 3.4): The server utilizes multi-dimensional aggregation parameters The aggregation process is completed according to the following formula: ， ,in, The aggregate weights for each edge core network, To update Global model at time step Indicates the number of communication rounds. Edge Core Network The communication model parameters; 3.5): The global model formed by the aggregation of the central core network. It is distributed to each edge core network.

7. The 6G edge network user prediction method based on fair federation according to claim 6, characterized in that, Step 4 specifically involves: using a 6G edge network user prediction mechanism based on fair federation. In this process, the training process iterates continuously between steps 2 and 3 until the maximum number of global communication rounds is reached. The Adam optimizer is used to train the communication models and personalized models of each edge core network.

8. The 6G edge network user prediction method based on fair federation according to claim 7, characterized in that, Step 5 specifically involves: the user trajectory data undergoing the 6G edge network user prediction mechanism based on fairness federation from step 4. After training, the communication models and local personalized models of each edge core network are deployed in its analysis logic function. The prediction results are pushed to the PCF through the Nnwdaf_AnalyticsInfo service. After receiving the prediction, the PCF issues mobility optimization strategies to the AMF. Subsequently, the AMF pre-interacts with the gNB to be connected through the N2 interface to achieve accurate analysis and real-time prediction of user data.

Images