A federated learning based distributed communication network optimization method

Abstract

The application discloses a distributed communication network optimization method based on federal learning, which comprises the following steps: at a client node, decomposing a local model update into a first update amount and a second update amount based on a preset decoupling rule, wherein the first update amount represents a general network rule, and the second update amount represents individualized demands of a specific network slice; uploading the two update amounts to a federal server by the client node; performing differential aggregation by the federal server, that is, aggregating the first update amounts of all clients to update a shared module of a federal model, and for each network slice type, only aggregating the second update amounts of the corresponding clients to update a corresponding individualized module; and the first update amount and the second update amount are separated by calculating factual gradients and counterfactual gradients, efficient individualized optimization of different network slices is realized under a federal learning framework, and the performance and generalization ability of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of communication optimization technology, and in particular to a distributed communication network optimization method based on federated learning. Background Technology

[0002] Network slicing, a core technology of next-generation communication networks, allows the creation of multiple virtual networks with different Quality of Service (QoS) requirements on the same physical network infrastructure. For example, enhanced mobile broadband (eMBB) slices require high peak rates, ultra-reliable low-latency communication (URLLC) slices require extremely low latency and high reliability, while massive machine-type communications (mMTC) slices need to support massive device connections. Federated learning (FL), as a distributed machine learning paradigm, allows for the collaborative training of a global model without sharing local raw data, making it well-suited for communication network optimization scenarios that protect operator data privacy. However, existing technologies face two fundamental contradictions when applying federated learning to network slicing optimization. First, data heterogeneity: the distribution of network state data (such as traffic load, channel quality, and interference levels) collected by network nodes (such as gNBs) in different slices and regions varies significantly. Second, model aggregation: standard federated learning aggregation algorithms (such as FedAvg) aim to learn an "average" model applicable to all clients. When data is highly heterogeneous, this aggregation method leads to a severe degrade in the performance of the global model, resulting in unsatisfactory optimization results for any specific slice, failing to meet its unique Quality of Service (QoS) requirements. Even by overlaying models or adjusting parameters, the problem cannot be fundamentally solved. Summary of the Invention

[0003] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0004] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a distributed communication network optimization method based on federated learning to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a distributed communication network optimization method based on federated learning, comprising:

[0006] On the client node, the model of local data is updated, and the model update is decomposed into a first update amount and a second update amount based on a preset decoupling rule;

[0007] The client node uploads the first update and the second update to the federated server;

[0008] The federation server performs differential aggregation on the first update and the second update received from one or more client nodes. The first update is aggregated to update the shared modules of the federation model, and the corresponding second update is aggregated for network slice types to update the personalized modules of the federation model.

[0009] As a preferred embodiment of the distributed communication network optimization method based on federated learning described in this invention, the first update quantity represents the general network rules of the network slice type, and the second update quantity represents the personalized requirements of the network slice type.

[0010] As a preferred embodiment of the distributed communication network optimization method based on federated learning described in this invention, the decoupling rule is a decoupling rule based on causal inference.

[0011] As a preferred embodiment of the distributed communication network optimization method based on federated learning described in this invention, the decoupling rule based on causal inference includes:

[0012] Based on the local real network data of the client node, the fact gradient is calculated as the initial amount for model update;

[0013] By intervening in the network slice causes in the real network data, counterfactual network data is generated, and counterfactual gradients are calculated based on the counterfactual network data;

[0014] The first update amount and the second update amount are determined based on the fact gradient and the counterfactual gradient.

[0015] As a preferred embodiment of the distributed communication network optimization method based on federated learning described in this invention, the first update amount is determined by the counterfactual gradient, and the second update amount is determined by the difference between the factual gradient and the counterfactual gradient.

[0016] As a preferred embodiment of the distributed communication network optimization method based on federated learning described in this invention, the federated model includes a shared backbone network and at least one personalized head network corresponding to a network slice type, wherein the shared module corresponds to the shared backbone network and the personalized module corresponds to the personalized head network.

[0017] As a preferred embodiment of the distributed communication network optimization method based on federated learning described in this invention, the differentiated aggregation includes:

[0018] The first update volume uploaded by all client nodes participating in federated learning is aggregated to update the shared backbone network;

[0019] For each network slice type, client nodes serving that slice type are selected, and only the second update quantity uploaded by these selected client nodes that corresponds to that slice type is aggregated to update the personalized header network.

[0020] The present invention also provides a federated server, comprising:

[0021] The update receiving module is used to receive the model update quantity uploaded by the client node, which has been decomposed into the first update quantity and the second update quantity;

[0022] The aggregation engine is configured to perform differential aggregation on the general updates and the personalized updates to update the shared modules and personalized modules in the federated model, respectively.

[0023] The present invention also provides a client node, which includes:

[0024] The update decomposition module is configured to decompose the local model update into a first update amount and a second update amount; wherein the general update is intended to reflect the general network rules applicable to multiple network slice types, and the personalized update is intended to reflect the personalized needs for a specific network slice type.

[0025] The update upload module is used to upload the general updates and the personalized updates to the federated server.

[0026] As a preferred embodiment of the client node described in this invention, the update decomposition module is configured to determine the first update amount and the second update amount by calculating the fact gradient and the counterfactual gradient.

[0027] Compared with existing technologies, the beneficial effects of this invention are as follows: By using a causal inference algorithm on the client side, this invention actively decouples model updates into general knowledge and personalized knowledge, and performs differentiated aggregation on the server side. Thus, under a unified federated learning framework, it achieves highly personalized optimization of different network slices. This not only fundamentally solves the problem of model performance degradation caused by data heterogeneity in traditional federated learning, but also improves the generalization ability and interpretability of the model. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0029] Figure 1 This is a flowchart illustrating the overall process of a distributed communication network optimization method based on federated learning, as described in one embodiment of the present invention. Detailed Implementation

[0030] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0031] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0032] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0033] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0034] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0035] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0036] Example 1

[0037] Reference Figure 1 This is the first embodiment of the present invention. This embodiment provides a distributed communication network optimization method based on federated learning, including: the optimization method runs in a distributed communication network system, the system including a central federated server and one or more client nodes;

[0038] It should be noted that the central federated server is a computing device deployed on the core side of the network, such as the network core network or the operator's cloud data center. The central federated server is configured to manage a federated model, which includes a shared backbone network module and personalized header network modules. The federated model is distributed to client nodes, and the server receives updates uploaded by the client nodes after applying decoupling rules based on causal inference. At the same time, a differentiated aggregation process is performed to update the federated model. For the client nodes, the client node is a computing entity at the network edge, such as a 5G base station, an edge computing server, or a user terminal with certain computing capabilities. Each client node is configured to receive and execute the federated model from the central federated server. This is achieved by collecting network status data within the client node's service range in real time, performing a gradient decoupling process based on causal inference decoupling rules on the locally trained model, and then uploading the decoupled local training model updates to the central federated server.

[0039] S1. On the client node, update the model of the local data and decompose the model update into a first update amount and a second update amount based on the preset decoupling rules;

[0040] Specifically, through the client node, network status data x and corresponding performance labels y are collected in real time within the service range of the client node to form a local dataset D;

[0041] Specifically, the performance label y is generated by the personalization header network in the personalization module;

[0042] Furthermore, to reflect the causal inference relationship, the network state data x is divided into two parts. The first part is the general network rules of network slice types, which are represented as the underlying network environment characteristics that are independent of network slice types.

[0043] Specifically, the characteristics of this underlying network environment include average reference signal received power, average reference signal received quality, estimated signal-to-noise ratio, reported channel quality indication, physical resource block utilization, and neighboring cell interference level.

[0044] Specifically, the average reference signal received power, such as the average reference signal received power of all active users in the serving cell over the past 100ms, reflects the overall signal coverage strength of the serving cell.

[0045] Specifically, the average reference signal reception quality can be expressed by the formula RSRQ = N × (RSRP / RSSI), where RSSI is a measurement value representing the total broadband power received by a user equipment (UE, such as a mobile phone) in a mobile communication network. It includes the signal power of the primary serving cell to be received, the interference signal power of all neighboring cells, and the power of background noise. N represents the number of resource blocks occupied by the RSSI measurement. RSRQ represents the reception quality of the reference signal. When the value of RSRQ is greater than the value of RSSI, it indicates that most of the reference signal in the total received power is a useful signal, and the interference and noise are very small. This indicates that the current situation is in an ideal environment with no load, single cell, or no interference. When the value of RSRQ is less than the value of RSSI, it indicates that the useful reference signal accounts for only a small part of the total received power, and most of it is strong interference from neighboring cells. This indicates that the current situation is in an area with high network load, cell edge, or multiple cell overlapping coverage.

[0046] Specifically, the estimated signal-to-noise ratio (SNR) represents the estimated value of the channel's SNR within the current scheduling period and determines the highest data rate that the base station can select for the user equipment. If the serving cell of the current user equipment changes, the estimated SNR will also change.

[0047] Specifically, the reported channel quality indication represents the quantization and reporting of the signal-to-noise ratio by the user equipment. The workflow is as follows: the user equipment measures the downlink reference signal from its serving cell, then the user equipment quantizes the signal-to-noise ratio by consulting the channel quality indication table defined by the 3GPP standard, and maps different signal-to-noise ratio ranges to integer values ​​from 0 to 15 according to the indication table. Finally, the user equipment sends this channel quality indication value to the base station of the serving cell through the uplink channel to complete the reporting.

[0048] Specifically, physical resource block utilization rate represents the proportion of allocated physical resource blocks to the total number of available physical resource blocks in the previous scheduling cycle, reflecting the load level of the serving cell.

[0049] Specifically, the neighboring cell interference level is obtained by estimating the total interference power from the downlink control channel of the neighboring cells;

[0050] Furthermore, the second part is the personalized requirements of network slices, which are represented as strong correlation features with the network slice type to which the data stream belongs, where each strong correlation feature is encoded using one-hot encoding.

[0051] Specifically, this strong correlation feature includes the following two cases:

[0052] For ultra-reliable low-latency communication (uRLLC), packet delay budget, packet error rate, and traffic patterns are considered;

[0053] For enhanced mobile broadband (eMBB), consider the service type (such as video streaming / file download), the current buffer fill (i.e., the number of bytes of data sent in the base station's buffer, reflecting the burstiness and urgency of the service) and the expected throughput (i.e., the expected rate).

[0054] It should be noted that for uRLLC and eMBB, the final performance label uses different decision functions in the output of the personalization module;

[0055] Specifically, the preset decoupling rule is a decoupling rule based on causal inference, and the calculation process of the decoupling rule is as follows:

[0056] Based on the local real network data of the client node, the fact gradient is calculated as the initial amount for model update;

[0057] It should be explained that the fact gradient is the gradient calculated on real, uninterrupted local data;

[0058] Specifically, the calculation of the fact gradient is as follows:

[0059] Initialize the local gradient accumulator;

[0060] Select data samples from local real network data and add them to the local dataset D;

[0061] Input the network state data x into the current local data model to obtain the prediction error of the current data, where the prediction error is obtained by the binary cross-entropy loss of the model;

[0062] The loss value of the current local data model is calculated using this error, and the gradient of the local data model is generated using the backpropagation method. This gradient is then used as the fact gradient.

[0063] It should be noted that the fact gradient is calculated by using the loss function to determine the partial derivative of the loss value with respect to the local data model parameters:

[0064]

[0065] Where G represents the loss function, and Θ represents the local data model parameters, such as weights and biases. Represented as the gradient of the local data model;

[0066] By intervening in the causes of network slicing in real network data, counterfactual network data is generated, and counterfactual gradients are calculated based on the counterfactual network data.

[0067] Specifically, by using a causal generative model, we can infer the potential general and specific causes from real data, intervene in the potential causes, and generate a potential code after intervention.

[0068] It should be explained that the counterfactual gradient is a gradient calculated on hypothetical data that has been causally manipulated;

[0069] Based on the fact gradient and the counterfactual gradient, determine the first update amount and the second update amount;

[0070] Specifically, keeping the underlying network environment features unchanged, the network slice features are replaced with a neutral, information-free baseline value based on the latent encoding. This baseline value can be the zero vector or the mean or median of the feature in all slice data, thus constructing counterfactual data.

[0071] Specifically, the gradient of the constructed counterfactual data and the original performance label y is calculated in the same way as the factual gradient above, that is, the partial derivative of the counterfactual data and the original performance label is calculated through the loss function to obtain the counterfactual gradient;

[0072] Furthermore, the first update amount is determined by the counterfactual gradient, and the second update amount is determined by the difference between the factual gradient and the counterfactual gradient.

[0073] Specifically, the counterfactual gradient is directly used as the first update amount to ensure that the gradient uploaded to the federated server for updating the shared backbone network is "cleaned" and pure general knowledge. The difference between the factual gradient and the counterfactual gradient is used as the second update amount because replacing the baseline value may produce a contaminated gradient, so the difference needs to be calculated.

[0074] S2. The client node uploads the first and second update values ​​to the federated server.

[0075] Specifically, each client node packages its first and second update portions (along with the slice type identifier of its service) and uploads them to the central federation server via a secure channel;

[0076] S3. Through the federation server, perform differential aggregation on the first update and the second update received from one or more client nodes. Aggregate the first update to update the shared module of the federation model, and aggregate the corresponding second update for the network slice type to update the personalized module of the federation model.

[0077] Specifically, the federated model consists of a shared module (shared backbone network) and one or more personalized modules (personalized head networks);

[0078] Furthermore, the initial update volume uploaded by all client nodes participating in federated learning is aggregated to update the shared backbone network;

[0079] Specifically, the shared backbone network employs a deep neural network with residual connections to prevent gradient vanishing and learn deeper features.

[0080] Specifically, the shared backbone network adopts a standard architecture of an input layer, hidden layers, and an input layer. The input layer dimension is the local dataset D. The hidden layer contains three residual blocks, each containing two fully connected (FC) layers. Each FC layer is followed by a ReLU activation function and a batch normalization (BN) layer. The structure of each residual block is as follows:

[0081] Block_1: FC (dimension of the input layer → 256) → BN → ReLU → FC (256 → 256) → BN;

[0082] Block_2: FC(256→128)→BN→ReLU→FC(128→128)→BN;

[0083] Block_3: FC(128→64)→BN→ReLU→FC(64→64)→BN;

[0084] In addition, the input of each residual block is added to the output of the next residual block through a skip connection;

[0085] Furthermore, for each network slice type, client nodes serving that slice type are selected, and only the second update volume uploaded by these selected client nodes that corresponds to that slice type is aggregated to update the personalized header network.

[0086] Specifically, the personalized head network uses a relatively shallow multilayer perceptron (MLP) to maintain the model's lightweight and professional nature, with the following architecture:

[0087] Input layer: Receives the shared feature vectors from the shared backbone network;

[0088] Hidden layers: consist of 1 to 2 fully connected layers, followed by an activation function, such as FC(64→32)→ReLU;

[0089] Output layer: The activation function selects the decision function based on the decision type output by the original performance label y;

[0090] Specifically, the central federated server calculates a weighted average of the first update volume uploaded by all clients and uses this average to update the shared backbone network;

[0091] Specifically, for each network slice type, the server selects client nodes serving that slice type and aggregates only the second update data uploaded by these selected client nodes that corresponds to that slice type. The aggregation process is as follows:

[0092] Specifically, the central federated server collects the first update data uploaded by all clients participating in this round of learning, performs a weighted average of the first update data to obtain the aggregation result, and updates the parameters of the shared backbone network of the central federated model.

[0093] Specifically, the polymerization process can be expressed by the following formula:

[0094]

[0095] in, This represents the new weight of the shared modules on the federated server at the start of round t+1 (i.e., the next round); This represents the old weights of the shared modules on the federated servers at the start of round t (i.e., the current round); the global learning rate of the central federated server controls the step size of each update operation and is set to 1; K represents the number of client nodes, i.e., how many clients there are. The weight of the client is represented by its data size n. k The proportion of data in the total data volume N determines the contribution of clients with larger data volumes to the federated model; Δw 1,k This is represented as the first update value;

[0096] Specifically, the central federation server groups the second update quantity based on the slice type identifier uploaded by the client. For the two network slice types mentioned above, the central federation server performs the following processing:

[0097] Filter out all clients that serve eMBB slices;

[0098] Only the second update volume uploaded by the selected clients, corresponding to the eMBB slice, is weighted and averaged.

[0099] Based on the weighted average of the first update, update the parameters of the eMBB personalized head network. Perform the same operation on uRLLC or other slice types to update their respective personalized head networks.

[0100] The above process is carried out iteratively until the federated model converges.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0102] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0106] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A distributed communication network optimization method based on federated learning, characterized in that, include: On the client node, the model of local data is updated, and the model update is decomposed into a first update amount and a second update amount based on a preset decoupling rule; The decoupling rule is a decoupling rule based on causal inference; The decoupling rules based on causal inference include: Based on the local real network data of the client node, the fact gradient is calculated as the initial amount for model update; By intervening in the network slice causes in the real network data, counterfactual network data is generated, and counterfactual gradients are calculated based on the counterfactual network data; The first update amount and the second update amount are determined based on the fact gradient and the counterfactual gradient; The first update amount is determined by the counterfactual gradient, and the second update amount is determined by the difference between the factual gradient and the counterfactual gradient; The client node uploads the first update and the second update to the federated server; The federation server performs differential aggregation on the first update and the second update received from one or more client nodes. The first update is aggregated to update the shared modules of the federation model, and the corresponding second update is aggregated for network slice types to update the personalized modules of the federation model.

2. The distributed communication network optimization method based on federated learning as described in claim 1, characterized in that, The first update quantity represents the general network pattern of the network slice type, and the second update quantity represents the personalized requirements of the network slice type.

3. The distributed communication network optimization method based on federated learning as described in claim 1, characterized in that, The federated model includes a shared backbone network and at least one personalized head network corresponding to a network slice type. The shared module corresponds to the shared backbone network, and the personalized module corresponds to the personalized head network.

4. The distributed communication network optimization method based on federated learning as described in claim 1 or 3, characterized in that, The differentiated aggregation includes: The first update volume uploaded by all client nodes participating in federated learning is aggregated to update the shared backbone network; For each network slice type, client nodes serving that slice type are selected, and only the second update quantity uploaded by these selected client nodes that corresponds to that slice type is aggregated to update the personalized header network.

5. A federated server, characterized in that, include: The update receiving module is used to receive the model update quantity uploaded by the client node, which has been decomposed into the first update quantity and the second update quantity; The first update amount is determined by the counterfactual gradient, and the second update amount is determined by the difference between the factual gradient and the counterfactual gradient; The fact gradient is obtained based on the local real network data of the client node; The counterfactual gradient intervenes based on the network slice cause in the local real network data to generate counterfactual network data, and is calculated based on the counterfactual network data; The aggregation engine is configured to perform differential aggregation on general updates and personalized updates to update the shared modules and personalized modules in the federated model, respectively. The first update amount is the general update, and the second update amount is the personalized update; The general updates are designed to reflect general network patterns applicable to multiple network slice types, while the personalized updates are designed to reflect personalized needs for specific network slice types. The execution of differentiated aggregation aims to aggregate the first update amount to update the shared modules of the federated model, and aggregate the corresponding second update amount for network slice type to update the personalized modules of the federated model.

6. A client node, characterized in that, include: The update decomposition module is configured to decompose local model updates into a first update amount and a second update amount. The decomposition is based on a preset decoupling rule, which is a decoupling rule based on causal inference. The decoupling rules based on causal inference include: Based on the local real network data of the client node, the fact gradient is calculated as the initial amount for model update; By intervening in the network slice causes in the real network data, counterfactual network data is generated, and counterfactual gradients are calculated based on the counterfactual network data; The first update amount and the second update amount are determined based on the fact gradient and the counterfactual gradient; The first update amount is determined by the counterfactual gradient, and the second update amount is determined by the difference between the factual gradient and the counterfactual gradient; The first update volume is a general update, and the second update volume is a personalized update; The general updates are designed to reflect general network patterns applicable to multiple network slice types, while the personalized updates are designed to reflect personalized needs for specific network slice types. The update upload module is used to upload the general updates and the personalized updates to the federated server.

7. The client node according to claim 6, characterized in that, The update decomposition module is configured to determine the first update amount and the second update amount by calculating the fact gradient and the counterfactual gradient.

Images