Longitudinal federated learning system construction method based on participant selection and parameter freezing

Abstract

The invention discloses a longitudinal federated learning system construction method based on participant selection and parameter freezing, and the method comprises the steps: evaluating and freezing a part of model parameters after the selection of a client, and reducing the communication traffic; the parameter server carries out aggregation according to the importance of the participants and sends an aggregated global model to the client; the client updates a local model and determines whether to freeze parameters according to the stability of the model; the participants of the next round of training are selected by evaluating hardware and statistical contributions of the participants; by freezing stable model parameters, the transmission quantity of intermediate statistical information is reduced, and the training efficiency is improved; on the premise of privacy protection, model convergence is accelerated, communication delay is reduced, and meanwhile model precision is kept.

Description

Technical Field

[0001] The present invention belongs to the field of computer network communication and machine learning technology, and relates to a method and device for constructing a vertical federated learning system, and specifically to a method and device for constructing a vertical federated learning system based on participant selection and parameter freezing. Background Art

[0002] Vertical Federated Learning (VFL) is a technique that allows geographically distributed data participants to collaboratively train shared machine learning models while protecting the privacy of their data. In VFL, each participant holds a different subset of features of the same instance. This approach is very useful in privacy-sensitive fields such as healthcare and e-commerce. As VFL scales up, training efficiency is limited by frequent long-distance communication from each participant to the central server. This communication is usually carried out over a wide area network (WAN), which brings latency and bandwidth challenges. In addition, some existing asynchronous VFL schemes speed up the training process, but may reduce the accuracy of the model due to model staleness. Therefore, how to improve the training efficiency and scalability of VFL while protecting privacy becomes a key issue.

[0003] In the current context, the main challenges facing vertical federated learning (VFL) include communication delays, model obsolescence, participant heterogeneity, computational overhead, and the balance between privacy protection and efficiency. Frequent communication over wide area networks leads to reduced training efficiency, while asynchronous schemes may lead to untimely model updates and affect accuracy. The data quality and hardware capabilities of the participants vary significantly, which may lead to unbalanced training. In addition, the complex calculations for evaluating the data contributions of participants increase the burden on the system, and it is necessary to effectively reduce the amount of communication and computational burden while protecting data privacy. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for constructing a longitudinal federated learning system by using Shapley values ​​to realize participant selection and adaptive parameter freezing, so as to solve the problems of communication delay, model obsolescence, participant heterogeneity, computational overhead, and the balance between privacy protection and efficiency.

[0005] The technical solution adopted by the method of the present invention is: a method for constructing a vertical federated learning system based on participant selection and parameter freezing, wherein the client local model of the vertical federated learning system is obtained through training, and the training process includes the following steps:

[0006] Step 1: At the beginning of each round of training, according to the statistical contribution value φ of each participant k (V) and processing time T k , evaluate the contribution of each participant and select the optimal set of participants;

[0007] Step 2: Each selected participant uses the local model θ k Generate embedding vectors And send it to the parameter server;

[0008] Step 3: The parameter server receives the embedding vectors of all participants Aggregate and calculate the loss function l(Θ), and the parameter server updates the global model parameter θ0;

[0009] Step 4: Check the change in local model parameters for each participant If the change is less than the set threshold ∈ s , then the update will be frozen according to the preset freezing period;

[0010] Step 5: At the end of the freeze period, recheck the local model parameter changes If the change exceeds the threshold ∈ s , then resume the update so that it can re-receive and apply the gradient update returned from the parameter server; otherwise, the system extends the length of the next freeze period;

[0011] Step 6: After every N rounds, re-evaluate the statistical contribution of the participants, adjust the participant selection strategy, and optimize the training efficiency; where N is a preset value;

[0012] Step 7: Repeat the above steps until the model reaches the preset accuracy or the maximum number of rounds is reached, and then the training ends.

[0013] Preferably, in step 1, by calculating the statistical contribution value of each participant To evaluate its statistical contribution, combined with the processing time T k To calculate the system utility Where T k is the processing time of the participant, T is the target round duration, and α is the penalty factor; the participant set is denoted as K = {P1, P2, …, P n}; all possible subsets of participants are denoted as S, excluding participant k; the utility of subset S is denoted as S(V); |S|! Calculate the weights of the permutations and combinations of participant k before and after being added to subset S. V(S) measures the loss reduction brought by the joint training model of participants in subset S alone, that is, the statistical contribution of subset S. {k} represents the set containing only participant k; represents the statistical contribution value of participant k in training round t; represents the indicator function, which is used to determine the processing time T of participant k k Whether it exceeds the target round length T; if T k >T, then apply the penalty factor α; use the online exploration-exploitation strategy by analyzing the historical performance of each participant Make adjustments Ensure that multiple participants are explored in the early stage and gradually converge to the optimal set of participants as training progresses, where ζ is the exploration parameter, t is the training round, and the number of times participant k is selected is recorded as n k ; Finally, the system is based on the system utility The participants are sorted and the participants whose total contribution reaches the set threshold ∈ are selected for training in the current round.

[0014] Preferably, the embedding vector is output for participants whose time exceeds a preset value and who are not selected Re-weighted, the weight is where τ k is the time since the participant last updated the model.

[0015] Preferably, in step 3, the server updates the global model parameter θ0 through a back propagation algorithm;

[0016] The specific implementation includes the following sub-steps:

[0017] Step 3.1: Aggregate and adjust the embedding vector;

[0018] The parameter server PS receives the t All embedding vectors of The embedding vector for each participant k is Applying a decay factor Adjusting the embedding vector To mitigate the impact of outdated models, where x i represents the i-th data sample, X t Represents the data sample set used in the tth round of training:

[0019] Aggregate the adjusted embedding vectors to form the global input data

[0020] Step 3.2: Forward propagation and loss calculation;

[0021] The aggregated embedding vector h agg Input into the global model to generate prediction results Calculate the comprehensive loss function l(Θ)=l1(Θ)+λ×l2(Θ);

[0022] Step 3.3: Back propagation and gradient calculation;

[0023] Back-propagate the comprehensive loss function l(Θ) to calculate the gradient of the global model parameter θ0

[0024] Step 3.4: Update global model parameters θ0;

[0025] Update global model parameters using optimization algorithms Where η is the learning rate;

[0026] Step 3.5: Gradient sending and local model update;

[0027] The calculated gradient Sent to unfrozen participants k∈P t , so that it can update its local model; each participant receives the gradient and uses backpropagation to update its local model parameters in is the gradient of the loss function for participant k;

[0028] Step 3.6: Dynamically adjust contribution weight and learning rate;

[0029] According to the contribution during the training process, the weight coefficient λ is dynamically adjusted to ensure that the model can effectively balance the prediction performance and contribution optimization at different training stages;

[0030] Dynamically adjust the learning rate based on the current round of training progress and model convergence status To accelerate convergence and avoid shock, where η0 is the initial learning rate, β is the decay coefficient, and t is the current round.

[0031] Preferably, in step 3, the loss function l(Θ)=l1(Θ)+λ×l2(Θ), where l1(Θ) is the prediction loss, Where N is the number of samples in the batch, C is the number of categories, and y i,c is the true label of the i-th sample in category c, is the model's predicted probability for the i-th sample in category c; l2(Θ) is the contribution adjustment loss function, Where P t represents the set of participants selected in the current round t, φ k (V) represents the contribution of participant k to model training, θ k represents the local model parameters of participant k, θ0 represents the global model parameters; λ is the weight coefficient used to balance the importance of prediction loss and contribution adjustment loss.

[0032] Preferably, in step 6, if a participant's Shapley value is higher than a preset value, the frequency of his / her participation in training is increased; otherwise, his / her selection frequency is reduced or he / she is temporarily excluded.

[0033] The present invention also provides a device for constructing a vertical federated learning system based on participant selection and parameter freezing, comprising:

[0034] one or more processors;

[0035] A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method for constructing a longitudinal federated learning system based on participant selection and parameter freezing.

[0036] The present invention also provides a product for constructing a longitudinal federated learning system based on participant selection and parameter freezing, including computer program instructions. When the computer program instructions are run on a computer, the computer executes the method for constructing a longitudinal federated learning system based on participant selection and parameter freezing.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] (1) The present invention proposes a participant selection mechanism based on Shapley values. Existing vertical federated learning (VFL) technologies usually adopt fixed or random selection strategies when selecting participants, which fails to fully consider the actual contribution of each participant to the model training process, resulting in high communication costs and low training efficiency. In addition, existing methods also fail to effectively deal with the computational bottleneck caused by differences in the hardware performance of participants. The present invention introduces a participant selection mechanism based on Shapley values, and for the first time realizes the accurate evaluation of the statistical contribution of participants in VFL, and combines the comprehensive consideration of hardware performance to dynamically select the most valuable set of participants for model training. By using the truncated Shapley method, the mechanism significantly reduces the complexity of Shapley value calculation, and can effectively resist the influence of lags and avoid training delays caused by differences in the hardware performance of participants. This not only greatly reduces the communication overhead in each round of training, but also improves the efficiency of the entire training process, ensuring that resource utilization and training time are optimized without compromising model accuracy.

[0039] (2) The adaptive parameter freezing strategy proposed in the present invention. In the existing vertical federated learning (VFL) technology, in each round of training, all participants will receive and apply the gradient updates returned from the parameter server, even though the model parameters of some participants have stabilized. This full-participation strategy leads to unnecessary communication overhead, prolongs the training time, and wastes computing resources. The present invention proposes an adaptive parameter freezing strategy. For those participants whose model parameters change less, the system will temporarily freeze their model update process and only retain the generation and transmission of embedding vectors. This strategy implements a dynamic freezing and recovery mechanism by real-time monitoring of parameter changes, ensuring that update operations are only performed when model parameters change significantly, thereby reducing unnecessary gradient transmission and computing burden. At the same time, the strategy also dynamically adjusts the length of the freezing period to adapt to the different stages of model parameter changes during training, minimizes communication and computing overhead, significantly improves training efficiency, and maintains the overall accuracy of the model. This adaptive parameter freezing strategy ensures the efficiency and reliability of the federated learning system while optimizing the utilization of communication and computing resources.

[0040] (3) The present invention proposes a dynamic adjustment of the participant selection strategy. The existing vertical federated learning (VFL) technology usually adopts a fixed participant selection strategy, which fails to effectively adapt to the dynamic changes in the contribution of each participant during the training process, resulting in the model being unable to fully utilize the data value of each participant at different stages, thereby affecting the overall training efficiency and model performance. The present invention proposes a dynamic adjustment of the participant selection strategy, which is based on re-evaluating the Shapley values ​​of all participants after several rounds, and reflects the actual contribution of the participants in the current training stage in real time. Through this dynamic adjustment mechanism, the system can flexibly adjust the frequency and priority of participant selection, so that the participants who contribute the most to the model update can be selected at each training stage. This strategy can not only automatically adapt to changes in participant data and system status, ensuring that the most valuable data and computing resources are always used during model training, but also prevent some participants from continuing to participate in training due to decreased contributions, thereby avoiding resource waste. Finally, the dynamic adjustment of the participant selection strategy significantly improves the training efficiency, optimizes resource allocation, and ensures the model convergence speed and final performance at different training stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The technical solution of this invention is further described below using embodiments and specific implementation methods. In addition, some drawings are also used in the process of describing the technical solution. For those skilled in the art, other drawings and the intention of the invention can also be obtained based on these drawings without paying creative labor.

[0042] Figure 1A deployment diagram of a vertical federation state learning system for participant selection and parameter freezing for an embodiment of the present invention;

[0043] Figure 2 A diagram for implementing a method for participant selection in an embodiment of the present invention;

[0044] Figure 3 This is an implementation diagram of the adaptive parameter freezing method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0046] The technologies involved in this embodiment include a participant selection mechanism based on Shapley values, an adaptive parameter freezing strategy, a dynamic participant selection adjustment mechanism, etc., which ultimately significantly improve the training efficiency and model performance of federated learning.

[0047] First, the formal definition of the t-th round symbol of this system is given:

[0048] Explore parameter ζ;

[0049] Penalty factor α;

[0050] Threshold ∈ s ,∈;

[0051] Participant set P = {P1, P2, ..., P n};

[0052] All possible subsets S of participants, excluding participant k;

[0053] The utility of subset S, the training loss function of the model is used to measure S(V);

[0054] The number of times participant k is selected n k ;

[0055] The statistical contribution value φ of participant k k (V);

[0056] Target round duration T;

[0057] Freeze period T freeze ;

[0058] The time T that participant k processes local data in this round of training k ;

[0059] The feature set of the i-th data sample held by participant k

[0060] The local model parameters θ of participant k k ;

[0061] Function to generate embedding vectors

[0062] Loss function l(Θ);

[0063] Model parameter changes

[0064] Please see Figure 1 This embodiment provides a method for constructing a vertical federated learning system based on participant selection and parameter freezing. The efficient vertical federated learning system (P2VFL) has a client local model obtained by training; the training process specifically includes the following steps:

[0065] Step 1: At the beginning of each round of training, select participants. Evaluate the contribution of each participant based on the Shapley value and hardware performance, and select the optimal set of participants;

[0066] Step 2: Each selected participant generates an embedding vector and sends it to the parameter server;

[0067] Step 3: The parameter server aggregates the received embedding vectors, calculates the loss and performs backpropagation to update the global model parameters;

[0068] Step 4: Check the model parameter changes of each participant. If the change is less than the set threshold, temporarily freeze the update to reduce unnecessary communication;

[0069] Step 5: At the end of the freeze period, recheck the model parameter changes and resume updating if necessary. Adjust the freeze period length as needed;

[0070] Step 6: After every few rounds, re-evaluate the participants’ Shapley values, adjust the participant selection strategy, and optimize the training efficiency;

[0071] Step 7: Repeat the above steps until the model reaches the preset accuracy or the maximum number of rounds is reached, and then the training ends.

[0072]

[0073]

[0074] Another specific implementation is also provided here. In order to improve the training efficiency of vertical federated learning (VFL), the paper proposes an algorithm based on participant selection such as Figure 2As shown in Figure 1, the optimal participant is dynamically selected by evaluating the system utility of each participant, thereby reducing communication overhead and accelerating model convergence. Combined with the statistical utility φ k (V) and hardware utility where φ k (V) The truncated Shapley value is used to estimate and quantify the contribution of participants to model training. In order to reduce the computational complexity, the algorithm truncates the Shapley value in each round of training, that is, when the marginal contribution of a participant exceeds the preset threshold ∈ s At the same time, in order to balance exploration and utilization, the algorithm introduces the upper confidence bound (UCB) strategy to adjust the system utility: in is a confidence term used to explore participants who are not frequently selected in the early stage. In addition, the algorithm also designs a contribution threshold ∈ to control the number of participants selected in each round, that is, when the sum of the statistical utilities of the selected participants reaches ∈ times the total utility, the selection is stopped: ∑ k∈P φ k (V)>∈∑ k∈K φ k (V) To address the problem of model staleness, the algorithm outputs embedding vectors for participants who have not been selected for a long time. Re-weighted, the weight is where τ k is the time since the last time the participant updated the model. The set of participants is denoted as K = {P1, P2, ..., P n}; all possible subsets of participants are denoted as S, excluding participant k; the utility of subset S is denoted as S(V); |S|! Calculate the weights of the permutations and combinations of participant k before and after being added to subset S. V(S) measures the loss reduction brought by the joint training model of participants in subset S alone, that is, the statistical contribution of subset S. {k} represents the set containing only participant k; represents the statistical contribution value of participant k in training round t; represents the indicator function, which is used to determine the processing time T of participant k k Whether it exceeds the target round length T; if T k >T, then the penalty factor α is applied;

[0075]

[0076]

[0077] Another specific implementation is provided here. In this embodiment, an adaptive parameter freezing strategy is proposed, such as Figure 3 As shown. This strategy monitors the stability of the participants' model parameters during the training process and dynamically freezes those parameters that have stabilized, thereby reducing communication overhead and accelerating model convergence. In VFL, since different parts of the model converge at different speeds during training, parameters close to the input layer usually stabilize earlier, while parameters close to the output layer require more updates. Therefore, the core idea of ​​adaptive parameter freezing is to temporarily freeze the update of these parameters when it is detected that the model parameters of some participants have stabilized, so as to reduce unnecessary gradient transmission. Specifically, the algorithm first defines a parameter stability judgment condition, that is, when the local model parameter θ of a certain participant k Gradient The change range is lower than a preset threshold ∈ s When , the parameter is considered to be stable and can be temporarily frozen: ||Δθ k ||<∈ s Once frozen, the participant’s model parameters θ k Will no longer be updated, and its corresponding embedding vector The gradients are no longer transmitted to the central parameter server (PS). This freezing mechanism can effectively reduce the gradient information that needs to be transmitted, thereby reducing the communication burden. In order to ensure the overall performance of the model, the algorithm also designs a thawing mechanism. Specifically, when a large error is detected in some parts of the global model, the frozen participant model will start updating again. This thawing operation is completed through a simple mechanism, that is, regularly checking the stability of the frozen parameters. If it is found that the model accuracy has decreased, the frozen state is unfrozen and these parameters continue to be updated. Through this adaptive parameter freezing mechanism, Algorithm 2 can significantly reduce communication overhead while maintaining model accuracy.

[0078] Another specific implementation is provided here, the back propagation algorithm, which includes the following sub-steps:

[0079] The specific implementation includes the following sub-steps:

[0080] Step 3.1: Aggregate and adjust the embedding vector;

[0081] Receiving embedding vectors: The parameter server (PS) receives embedding vectors from the selected set of participants P t All embedding vectors of

[0082] Apply the decay factor: to the embedding vector of each participant k Applying a decay factor Adjust the embedding vector to mitigate the impact of the stale model, where x i represents the i-th data sample, X t Represents the data sample set used in the tth round of training:

[0083]

[0084] Aggregate embedding vectors: Aggregate the adjusted embedding vectors to form a global input vector h agg , by averaging

[0085]

[0086] Step 3.2: Forward propagation and loss calculation;

[0087] Forward propagation: The aggregated embedding vector h agg Input into the global model to generate prediction results Calculate the comprehensive loss function l(Θ)=l1(Θ)+λ×l2(Θ);

[0088] Step 3.3: Back propagation and gradient calculation;

[0089] Back propagation: Back propagate the comprehensive loss function l(Θ) to calculate the gradient of the global model parameter θ0

[0090]

[0091] Step 3.4: Update global model parameters θ0;

[0092] Gradient descent update: Update the global model parameters θ0 using an optimization algorithm such as stochastic gradient descent (SGD), where η is the learning rate.

[0093]

[0094] Step 3.5: Gradient sending and local model update;

[0095] Send gradient: Send the calculated gradient Sent to unfrozen participants k∈P t , so that it updates the local model.

[0096] Local update: Each participant receives the gradient and updates its local model parameters θ using backpropagation k ,in is the gradient of the loss function for participant k.

[0097]

[0098] Step 3.6: Dynamically adjust contribution weight and learning rate;

[0099] Dynamic adjustment of contribution weight: Dynamically adjust the weight coefficient λ according to the contribution during the training process to ensure that the model can effectively balance prediction performance and contribution optimization at different training stages.

[0100] Learning rate adjustment: Based on the training progress and model convergence status of the current round, the learning rate η is dynamically adjusted to accelerate convergence and avoid oscillation, where η0 is the initial learning rate, β is the decay coefficient, and t is the current round:

[0101]

[0102] Another specific implementation is provided herein, wherein the loss function l(Θ)=l1(Θ)+λ×l2(Θ), wherein l1(Θ) is the prediction loss, Where N is the number of samples in the batch, C is the number of categories, and y i,c is the true label of the i-th sample in category c, is the model's predicted probability for the i-th sample in category c; l2(Θ) is the contribution adjustment loss function, which is used to adjust the update of model parameters to ensure that high-contribution participants have a greater influence on the model and that the influence of low-contribution participants is appropriately reduced, which helps optimize training efficiency and maintain the fairness and accuracy of the model; Where P t represents the set of participants selected in the current round t, φ k (V) represents the contribution of participant k to model training, θ k represents the local model parameters of participant k, θ0 represents the global model parameters; λ is the weight coefficient used to balance the importance of prediction loss and contribution adjustment loss.

[0103] This embodiment also provides a device for constructing a vertical federated learning system based on participant selection and parameter freezing, including:

[0104] one or more processors;

[0105] A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method for constructing a longitudinal federated learning system based on participant selection and parameter freezing.

[0106] This embodiment also provides a product for constructing a longitudinal federated learning system based on participant selection and parameter freezing, including computer program instructions. When the computer program instructions are executed on a computer, the computer executes the method for constructing a longitudinal federated learning system based on participant selection and parameter freezing.

[0107] The present invention proposes an efficient vertical federated learning framework, which aims to improve the training efficiency of federated learning and reduce communication overhead by optimizing participant selection and model parameter update strategies. In order to cope with the contribution differences between participants and the problem of uneven hardware performance, the present invention introduces a participant selection mechanism based on statistical contribution value (Shapley value), dynamically evaluates the statistical contribution of participants, and selects the optimal set of participants to participate in training in combination with hardware performance indicators. In addition, in order to further reduce unnecessary communication and computing burdens, the present invention proposes an adaptive parameter freezing strategy, which temporarily freezes the model update of participants with small parameter changes, and only retains the generation and transmission of embedded vectors, thereby reducing communication overhead. In order to cope with the dynamic changes in participant contribution during training, the present invention also designs a dynamic adjustment participant selection strategy, which adjusts the participant selection frequency in real time by regularly re-evaluating the Shapley value to ensure optimal resource utilization during training.

[0108] It should be understood that the embodiments described above are part of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be combined with each other arbitrarily to form a feasible technical solution. Such combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0109] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A method for constructing a vertical federated learning system based on participant selection and parameter freezing, characterized by: In the vertical federated learning system, the client local model is obtained through training, and the training process includes the following steps: Step 1: At the beginning of each round of training, according to the statistical contribution value φ of each participant k (V) and processing time T k , evaluate the contribution of each participant and select the optimal set of participants; Step 2: Each selected participant uses the local model θ k Generate embedding vectors And send it to the parameter server; Step 3: The parameter server receives the embedding vectors of all participants Aggregate and calculate the loss function l(Θ), and the parameter server updates the global model parameter θ0; Step 4: Check the change in local model parameters for each participant If the change is less than the set threshold ∈ s , then the update will be frozen according to the preset freezing period; Step 5: At the end of the freeze period, recheck the local model parameter changes If the change exceeds the threshold ∈ s , then resume the update so that it can re-receive and apply the gradient update returned from the parameter server; otherwise, the system extends the length of the next freeze period; Step 6: After every N rounds, re-evaluate the statistical contribution of the participants, adjust the participant selection strategy, and optimize the training efficiency; where N is a preset value; Step 7: Repeat the above steps until the model reaches the preset accuracy or the maximum number of rounds is reached, and then the training ends.

2. The method for constructing a vertical federated learning system based on participant selection and parameter freezing according to claim 1, characterized in that: In step 1, by calculating the statistical contribution value of each participant To evaluate its statistical contribution, combined with the processing time T k To calculate the system utility Where T k is the processing time of the participant, T is the target round duration, and α is the penalty factor; the participant set is denoted as K = {P1, P2, …, P n }; all possible subsets of participants are denoted as S, excluding participant k; the utility of subset S is denoted as S(V); |S|! Calculate the weights of the permutations and combinations of participant k before and after being added to subset S. V(S) measures the loss reduction brought by the joint training model of participants in subset S alone, that is, the statistical contribution of subset S. {k} represents the set containing only participant k; represents the statistical contribution value of participant k in training round t; represents the indicator function, which is used to determine the processing time T of participant k k Whether it exceeds the target round length T; if T k >T, then apply the penalty factor α; use the online exploration-exploitation strategy by analyzing the historical performance of each participant Make adjustments Ensure that multiple participants are explored in the early stage and gradually converge to the optimal set of participants as training progresses, where ζ is the exploration parameter, t is the training round, and the number of times participant k is selected is recorded as n k ; Finally, the system is based on the system utility The participants are sorted and the participants whose total contribution reaches the set threshold ∈ are selected for training in the current round.

3. The method for constructing a vertical federated learning system based on participant selection and parameter freezing according to claim 2, characterized in that: Output embedding vector for participants who have not been selected for more than a preset time Re-weighted, the weight is where τ k is the time since the participant last updated the model.

4. The method for constructing a vertical federated learning system based on participant selection and parameter freezing according to claim 1, characterized in that: In step 3, the server updates the global model parameter θ0 through the back-propagation algorithm; The specific implementation includes the following sub-steps: Step 3.1: Aggregate and adjust the embedding vector; The parameter server PS receives the t All embedding vectors of The embedding vector for each participant k is Applying a decay factor Adjusting the embedding vector To mitigate the impact of outdated models, where x i represents the i-th data sample, X t Represents the data sample set used in the tth round of training: Aggregate the adjusted embedding vectors to form the global input data Step 3.2: Forward propagation and loss calculation; The aggregated embedding vector h agg Input into the global model to generate prediction results Calculate the comprehensive loss function l(Θ)=l1(Θ)+λ×l2(Θ); Step 3.3: Back propagation and gradient calculation; Back-propagate the comprehensive loss function l(Θ) to calculate the gradient of the global model parameter θ0 Step 3.4: Update global model parameters θ0; Update global model parameters using optimization algorithms Where η is the learning rate; Step 3.5: Gradient sending and local model update; The calculated gradient Sent to unfrozen participants k∈P t , so that it can update its local model; each participant receives the gradient and uses backpropagation to update its local model parameters in is the gradient of the loss function for participant k; Step 3.6: Dynamically adjust contribution weight and learning rate; According to the contribution during the training process, the weight coefficient λ is dynamically adjusted to ensure that the model can effectively balance the prediction performance and contribution optimization at different training stages; Dynamically adjust the learning rate based on the current round of training progress and model convergence status To accelerate convergence and avoid shock, where η0 is the initial learning rate, β is the decay coefficient, and t is the current round.

5. The method for constructing a vertical federated learning system based on participant selection and parameter freezing according to claim 1 or 4, characterized in that: In step 3, the loss function l(Θ)=l1(Θ)+λ×l2(Θ), where l1(Θ) is the prediction loss, Where N is the number of samples in the batch, C is the number of categories, and y i,c is the true label of the i-th sample in category c, is the model's predicted probability for the i-th sample in category c; l2(Θ) is the contribution adjustment loss function, Where P t represents the set of participants selected in the current round t, φ k (V) represents the contribution of participant k to model training, θ k represents the local model parameters of participant k, θ0 represents the global model parameters; λ is the weight coefficient used to balance the importance of prediction loss and contribution adjustment loss.

6. The method for constructing a vertical federated learning system based on participant selection and parameter freezing according to claim 1, characterized in that: In step 6, if a participant's Shapley value is higher than a preset value, the frequency of his / her participation in training is increased; otherwise, his / her selection frequency is reduced or temporarily excluded.

7. A device for constructing a vertical federated learning system based on participant selection and parameter freezing, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method for constructing a longitudinal federated learning system based on participant selection and parameter freezing as described in any one of claims 1 to 5.

8. A product for building a vertical federated learning system based on participant selection and parameter freezing, comprising computer program instructions, characterized in that: When the computer program instructions are executed on a computer, the computer is caused to execute the method for constructing a longitudinal federated learning system based on participant selection and parameter freezing as described in any one of claims 1 to 6.

Images