A cloud desktop throughput prediction method for cold start and mode forgetting

Abstract

The application discloses a cloud desktop throughput prediction method for cold start and mode forgetting, comprising the following steps: step 1, similarity attention federated learning assisted training; step 2, throughput prediction for safe cold start in the initial stage of a new user; step 3, fluctuation-residual feature extraction based on a TDLinear network model; step 4, model updating based on an adaptive memory module and EWC; and step 5, throughput prediction for training forgetting in the long-term use stage of a user. Through a federated learning algorithm, the existing throughput historical use knowledge of other users is utilized to assist training under the premise of guaranteeing user data privacy. Old modes are continuously reviewed based on continuous incremental prediction, and a memory module is designed to incrementally update new modes, so that the occurrence of forgetting in training is avoided, and the advantages of two stages are combined to realize the throughput prediction of a cloud desktop user in the whole life cycle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video stream throughput prediction in cloud desktop scenarios, in particular to a cloud desktop throughput prediction method for cold start and mode forgetting. BACKGROUND

[0002] At present, with the development of cloud computing, remote virtual desktop services built on the cloud are attracting attention and are widely used in education, office, government and other scenarios. Compared with traditional PCs, cloud desktops are managed by back-end operation and maintenance personnel, avoiding hardware and software failures and frequent upgrades. At the same time, cloud desktops have higher flexibility, can support multi-terminal access anytime and anywhere, and can meet various scene requirements by using different configurations. However, the development of cloud desktops is always affected by bandwidth. The popularity of gigabit networks has solved the problem of cloud desktop use, such as lag and low resolution, but the problem of excessive allocation of bandwidth still exists. It is understood that most network bandwidth resource allocation currently uses a fixed strategy, which not only causes resource waste but also brings high costs to users, so a throughput prediction scheme for guiding cloud desktop network bandwidth resource allocation is urgently needed.

[0003] Existing mature cloud desktop products and research mainly consider how to optimize coding and decoding to improve user experience, and pay little attention to network problems related to throughput and bandwidth resource allocation. Network bandwidth resources are the key factor in determining the smoothness and stability of cloud desktop services. Fixed bandwidth cannot meet the multi-scene needs of users, and inevitably causes resource waste in under-saturation and network transmission limitations in saturation. A real-time throughput prediction scheme helps guide dynamic bandwidth allocation. Cloud desktop bandwidth prediction based on historical data mainly faces two problems of "cold start" and "training forgetting". When a new user first uses the service, there is little historical data to support model training to achieve accurate throughput prediction, and there is a "cold start" problem. A feasible solution is to combine existing knowledge or recommend strategies. However, the difficulty in the cloud desktop scenario is that while using the knowledge of other users, data isolation between them needs to be ensured, and "safe start" is achieved under the premise of ensuring data privacy.

[0004] In addition, as the user's use time increases, historical data accumulates, and new knowledge will cover part of the old knowledge in the long-term training process, causing the "training forgetting" problem. Constantly reviewing old knowledge is a feasible solution. There are frequent changes in use scenarios during the use of the user's cloud desktop. Another big problem is how to effectively avoid the coverage of old scenarios by new scenario knowledge learned in the training process, and effectively use the time sequence features.

[0005] To sum up, there are at least the following technical problems:

[0006] The cloud desktop bandwidth prediction based on historical data mainly faces two problems of "cold start" and "training forgetting".

[0007] With the increase of user usage time, historical data accumulates a lot, and new knowledge learned in the long-term training process will cover part of the old knowledge, resulting in the problem of "training forgetting".

[0008] How to effectively avoid the coverage of new scene knowledge learned in the training process to the old scene, and effectively use the time sequence characteristics. SUMMARY

[0009] The main purpose of the present application is to provide a cloud desktop throughput prediction method for cold start and mode forgetting, to solve the problem that the cloud desktop bandwidth prediction based on historical data in the prior art mainly faces two problems of "cold start" and "training forgetting". With the increase of user usage time, historical data accumulates a lot, and new knowledge learned in the long-term training process will cover part of the old knowledge, resulting in the problem of "training forgetting". How to effectively avoid the coverage of new scene knowledge learned in the training process to the old scene, and effectively use the time sequence characteristics.

[0010] In order to achieve the above purpose, according to one aspect of the present application, a cloud desktop throughput prediction method for cold start and mode forgetting is provided, comprising:

[0011] Step 1: similarity attention federated learning assisted training;

[0012] Step 2: throughput prediction for initial stage of new user safe cold start;

[0013] Step 3: fluctuation-residual feature extraction based on TDLinear network model;

[0014] Step 4: model update based on adaptive memory module and EWC;

[0015] Step 5: throughput prediction for training forgetting in long-term use stage of user.

[0016] Preferably, in step 1, the overall framework of horizontal federated learning is adopted, which includes global side and local user side, the global side is used to process the data uploaded by each user, and the result is returned to each local user after common processing.

[0017] Preferably, in step 1, the prior knowledge of new users is first obtained, the prior knowledge is generated when the user registers or binds an account, including basic identity information and use information.

[0018] Preferably, in step 1, the user's prior knowledge is compared to select a user group with high relevance as a federal candidate set for a new user, and in this process, the user uploads the respective data to the global side for screening operation, and the same lightweight prediction model multi-layer perception is used on the global side and the local user side, the federal candidate set is trained separately and the gradient parameters in the training process are uploaded to the global side, while the new user also trains with a small amount of existing data and uploads the training parameters.

[0019] Preferably, in step 1, the global side calculates the similarity of the obtained training parameters, and through two rounds of interaction with the federal reserve users, the similarity weight is obtained, and the updated training parameters are calculated based on the weight and returned to the new user to update the model, in this process, the new user only interacts with the global side once, and needs to wait for the interaction between the federal candidate set and the global side.

[0020] Preferably, in step 2, the training loss before and after the local end obtains two training updates is uploaded to the global side to calculate the similarity between the new user and the candidate user, and the more the loss function obtained by the global side aggregation parameter training is reduced compared with the original loss function, the higher the similarity between the candidate user and the new user, and it is considered that the training parameters of the user have greater guiding significance, on the contrary, when the loss function is reduced or even increased, it is considered that the similarity is lower, and the guiding significance of the training parameters of the user is smaller, from which the m users with the highest similarity are selected to assist the new user to complete the throughput prediction task, considering the real-time requirement of prediction, in the initial stage, the basic multi-layer perception model is used for throughput prediction, the attention weight of the candidate user's parameters is set according to the similarity, the weight of the user with greater guiding significance is higher, and the prediction is completed through weighting.

[0021] Preferably, in step 3, as the use time increases, the historical data of the new user gradually accumulates, and when it exceeds a certain threshold, it can complete the training independently, at this time, the TDLinear model is designed to extract the features of the fluctuation part and the stable part of the historical data of the throughput, when the user is more active in using the cloud desktop, the throughput presents a sharp fluctuation, and when the cloud desktop is idle, the throughput is relatively stable, a sliding time window is divided, the variance in the window is calculated to measure the fluctuation degree of the throughput, so that the two features are extracted respectively, and the memory module and EWC are used to review the historical information for the two features.

[0022] Preferably, in step 4, based on the extracted fluctuation part characteristics, the old mode is stored and the new mode is updated through the adaptive memory module, the input data is passed through the full connection layer, the characteristics are retained on a small number of neurons, the memory module is used to store the characteristic neurons of different modes, the Pearson correlation coefficient is used to measure the correlation between the characteristic neurons in the current window and the neurons in the memory module, if the correlation is high, the memory module neurons are directly used to complete the training, if the correlation is low, the memory module is stored to complete the update, for stable part characteristics, the model parameters of the sliding window are constrained, and the same penalty term is used to review the past training parameters to constantly review the old knowledge.

[0023] Preferably, in step 5, after the division of the TDLinear model characteristics, the fluctuation characteristics and the stable characteristics of the idle time of the user active use of the cloud desktop are extracted respectively, the fluctuation characteristics are updated adaptively by continuously measuring the correlation of the neurons through the memory module, the network generalization ability is improved through Dropout regularization to prevent model overfitting, Ensemble is used to balance the prediction results of different memory units, and finally a linear prediction model is used to obtain the first prediction result y act .

[0024] Preferably, for stable characteristics, EWC is used to review historical training parameters, in order to retain the knowledge of previous tasks, key parameters are punished so that they will not be easily changed when training new tasks, and the same linear model is used to obtain the second prediction result y idle The two parts are aggregated in the time dimension to obtain the final output result y , wherein y act is the first prediction result; y idle is the second prediction result; and y pred is the final output result.

[0025] The technical scheme of the application has the following technical effects:

[0026] Through the federated learning algorithm, the existing throughput historical use knowledge of other users is used to assist the training under the premise of ensuring user data privacy. With long-time use, when the data volume accumulates enough, the user will independently use the historical data to complete the training. The present application proposes to continuously review the old mode based on continuous incremental prediction, and designs a memory module to incrementally update the new mode, which has the technical effect of avoiding forgetting during training. The two stages are combined to realize the throughput prediction of the whole life cycle of the cloud desktop user. BRIEF DESCRIPTION OF DRAWINGS

[0027] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0028] Figure 1 A total flow chart of a cold start and mode forgetting oriented cloud desktop throughput prediction method according to the present application is shown;

[0029] Figure 2 A similarity attention based federated learning framework diagram adopted by the cold start and mode forgetting oriented cloud desktop throughput prediction method in Figure 1

[0030] Figure 3 A resilient weight keeping incremental learning model diagram adopted by the cold start and mode forgetting oriented cloud desktop throughput prediction method in Figure 1

[0031] Figure 4 A TDLinear based fluctuation-residual feature extraction model diagram adopted by the cold start and mode forgetting oriented cloud desktop throughput prediction method in Figure 1

[0032] Figure 5 A multi-layer perceptron structure model diagram of the cold start and mode forgetting oriented cloud desktop throughput prediction method in Figure 1 DETAILED DESCRIPTION

[0033] It should be noted that the embodiments and the features in the embodiments in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0034] As shown in Figures 1 to 5 The present application provides a cold start and mode forgetting oriented cloud desktop throughput prediction method, which comprises the following steps: step 1: similarity attention based federated learning assisted training; step 2: throughput prediction oriented to safe cold start in the initial stage of new users; step 3: fluctuation-residual feature extraction based on a TDLinear network model; step 4: model updating based on an adaptive memory module and EWC; and step 5: throughput prediction oriented to training forgetting in the long-term use stage of users.

[0035] ​​​​This embodiment comprehensively considers the entire lifecycle of a user's cloud desktop usage. At the initial stage of use, when the amount of new user data is insufficient and independent prediction is difficult, this embodiment designs a federated learning algorithm that, while ensuring user data privacy, leverages the historical throughput usage knowledge of other users to assist in training. With extended use, when a sufficient amount of data has accumulated, users will independently complete training using their own historical data. This solution proposes to continuously review old patterns based on continuous incremental prediction, designing a memory module to incrementally update new patterns to avoid forgetting during training. A two-stage approach is used to achieve throughput prediction for cloud desktop users throughout their entire lifecycle. The following details the two-stage prediction of this solution.

[0036] Phase 1: New User Safe Cold Start Solution

[0037] 1.1 Auxiliary Training Strategy Based on Similarity Attention Federated Learning

[0038] In cloud desktop scenarios, each user uses the cloud desktop as an independent entity to complete their own tasks, maintaining confidentiality and maintaining a separate connection. When a new user begins using the cloud desktop, accurate throughput prediction cannot be achieved through independent training using only their limited historical data. Historical usage patterns of similar users can be instructive. To leverage the general patterns of historical throughput usage among other users while ensuring user data security, this solution employs a federated learning framework to address the "secure cold start" problem. Furthermore, an attention mechanism is designed based on user similarity to adjust the guidance weights for different users.

[0039] Considering that the application scenarios here are similar, that is, different users use cloud desktops to complete business, the present invention adopts the overall framework of horizontal federated learning, which is mainly composed of a global side and a local side. The global side is used to process the data uploaded by each user, and the results are returned to each local user after joint processing. First, get the new user user new Prior knowledge of user prior The prior knowledge is generated when the user registers or binds an account, including but not limited to the user's occupation, cloud desktop purpose and other basic identity information and usage information. By comparing the user's prior knowledge, the top-m user groups with high relevance are selected for the new user. U={u1,u2,…,u m}as a federal candidate set. In this process, users upload their respective data to the global side for encryption, and the screening operation is completed by the global side to protect individual privacy. To ensure the speed of model training and the real-time of prediction, the same lightweight prediction model multilayer perceptron is used on the global side and the local user side. The federal candidate set U will be trained separately and upload the gradient and other parameters during the training process to the global side for encryption, while new users also train with a small amount of existing data, and upload the training parameters for encryption. The global side will calculate the similarity of the obtained training parameters, and through two rounds of interaction with the federal reserve users U, the similarity weight Similarity = {s1, s2,..., sn} is obtained m}Based on the weight, the updated training parameters are calculated and returned to the new user user new Update the model. In this process, the new user only interacts with the global side once, and needs to wait for the interaction between the federal candidate set U and the global side.

[0040] 1.2 Model parameter similarity attention weight calculation

[0041] In this federated learning algorithm, how to use the training parameters provided by the federal candidate set U to guide the new user user new to complete the update and training of the parameters is the key part. According to the influence of different parameters on the gradient descent of the training loss function, the similarity is defined, and through two rounds of iteration, the similarity weight set of the federal reserve U and the new user can be calculated, and the weight set is used to update the training parameters to guide the training of the new user. The specific calculation process is as follows:

[0042] For a new user user new , define its training model M new The parameters θ new and loss L new are generated after training, and the training model set of the federal reserve user U is The parameter set generated by training is and the loss The parameter θ i is aggregated after the global model M global The parameter is First, the global side aggregates the parameters in the model:

[0043]

[0044] Among them, is the updated parameter after training; M global is the global model; θ new is the parameter generated after the training of the new user; is the training model set of the federal reserve user U The parameter set generated by training.

[0045] Here, the parameters of the new user are aggregated with the parameters in the Federal Reserve respectively, and then trained to obtain the updated parameters after training These parameters are returned to the local for secondary training, and the updated loss after training Again, to the global side, calculate the similarity:

[0046]

[0047] Where, s i represents the similarity of user i in the federal candidate set to the new user, and the similarity set Similarity={s1,s2,…,s m}; is a normalization operation, and the result is normalized to the range [-1, 1]; L i and L i represent the original and updated losses of user i, respectively, M i and M global represent the training model of the i-th user and the global model, respectively, θ i and θ new represent the training parameters of user i and the parameters generated by the new user after training, respectively; B is the batch size of training, and at this time, the training parameters of the new user after updating according to the similarity attention are:

[0048]

[0049] Where, represents the training parameters updated using the similarity attention, λ is the weight of the parameters of the new user after aggregation, θ new is the parameter generated by the training of the new user, s i is the similarity of user i in the federal candidate set to the new user, is the training parameter of user i on the server side; in the similarity calculation scheme proposed, the more the loss function trained by the global side after aggregation is reduced compared with the original, the higher the similarity of the user to the new user, and the scheme will consider that the training parameter of the user has greater guiding significance, and vice versa. When the loss is reduced less or even increased, it is considered that the guiding significance of the training parameter of the user is smaller. At this time, the weight is set to 0. Through the above iteration process, the parameters of the new user are updated in combination with the federal candidate user set U.

[0050] 2. Stage two: training forgetting solution based on continuous incremental learning

[0051] When the user uses for a certain period of time, the historical bandwidth usage record will increase, and when the data volume accumulates more than the threshold τ, the historical throughput data of the user is no longer combined with the data of the other users, and the prediction is completed independently by using the historical throughput data. Considering that new patterns will have a certain covering effect on old patterns during training, the scheme proposes a continuous incremental learning model to complete training and prediction. The input throughput data is divided into fluctuation part and remaining part. The fluctuation part belongs to the part where the user is active in using the cloud desktop, and the remaining part is considered to be the idle state of the user. The two parts are extracted for prediction, and finally the prediction results are obtained by splicing according to time. The whole process uses an adaptive incremental model to complete the update.

[0052] 2.1 Fluctuation-remaining feature extraction based on TDLinear network model

[0053] First, the feature extraction rules of the fluctuation part and the remaining part of the input data in the scheme will be introduced. The preprocessing of the input data is carried out in a sliding window manner. Before training, the data is divided into fluctuation part and remaining part according to the use state of the user's cloud desktop. Through data cleaning, it is found that when the user uses the cloud desktop to complete related tasks, the bandwidth will produce relatively violent fluctuations, and when the cloud desktop is idle, the bandwidth usage is stable and remains at a low state. Therefore, through training, the model learns all the throughput fluctuation patterns and general stable patterns of the user, that is, the throughput feature rules of all services of the user when using the cloud desktop. Specifically, the scheme calculates the variance of the data in a sliding window. The part greater than or equal to the variance threshold is classified as the feature of the user actively using the cloud desktop, and the part less than it is considered to be the general rule of the user using the service. After dividing the input data into two parts, the data is extracted through the fully connected layer, and the number of input neurons is reduced through the fully connected layer to retain the features on as few neurons as possible. In order to save the patterns under different windows, the scheme designs a memory module to store the feature neurons of different patterns. The Pearson correlation coefficient is used to measure the similarity between the feature neurons in the current window and the neurons in the memory module:

[0054]

[0055] Where r xy is the Pearson coefficient, n is the number of neurons, x i is the i-th neuron in the current window, and y iThe i-th neuron in the memory module is recorded. When the feature neurons in the sliding window are similar to the existing ones in the memory module, the training of the sliding window is directly skipped, the sliding window is moved forward, and the memory module is directly used for training. When the feature neurons have never appeared in the memory module, they are stored in the memory module and further embedding operation is performed; for the remaining part (stable part), further operation is completed using elastic weight retention (EWC).

[0056] 2.2 Incremental continual learning based on EWC

[0057] For fluctuating throughput data, when the data in the sliding window is considered to have new patterns by the memory module, the embedding method is needed to aggregate the feature vectors of the new patterns with the original feature vectors for the learning process of the memory module. The present application stores the data in the sliding window as new data in the memory module, and splices the new feature vectors after the same proportion of Dropout with the original feature vectors, ensuring that the proportion of feature vectors retained by each window is the same. For the remaining data, the processing method similar to the fluctuating feature is adopted, except that considering that the training difference of the remaining part should not be too large, the present application constrains the model parameters of the updated sliding window each time, using the EWC strategy. The core idea of this strategy is: when learning a new task, calculate the importance of each parameter in the previous training process, and assign a weight to each parameter according to the importance; at the same time, the parameter is also constrained when learning a new task, so that it will not change too much, thereby achieving the purpose of preserving old knowledge. Specifically, the importance weight of the parameter is generally obtained by calculating its Fisher matrix, which is widely used in statistics and machine learning fields to measure the uncertainty of parameter estimation. At this time, the importance weight of the parameter θ i can be calculated by the following formula:

[0058]

[0059] Where F i,t is the Fisher information matrix of the parameter θ i about task t. EWC also constrains the elastic weight of the new task parameter through a regularization term:

[0060]

[0061] Where B is the new task and A is the old task. L B (θ) represents the loss function of directly training the new task without adding EWC, F i represents the Fisher information matrix, and θ iθold A,i θold

[0062] Specifically, the method mainly includes the following steps:

[0063] S1: an auxiliary training strategy based on similarity attention federated learning

[0064] The strategy is mainly composed of a local user end and a global side. During training, the local user only uploads the model training parameters to the global side, ensuring the safety and privacy of the data. For a new user user new , when registering for cloud desktop services, it often contains some basic information user prior . By comparing this information, a candidate set of federated users U with similar information to the new user is determined. On the local side, the new user and the candidate users U first use a common network structure and use their own data to perform the first round of training, producing training loss and training parameters The model training parameters are uploaded to the global side together. On the global side, the parameters of U are aggregated one by one with the new user, and the second round of training is completed together to update the model parameters and return them to the local end. The local end uses the updated parameters to complete the third round of training, producing updated loss The influence of the parameters of the users in the candidate set is measured by comparing the changes in the loss function.

[0065] S2: a throughput prediction scheme for safe cold start in the initial use stage of a new user

[0066] The training losses before and after the two training updates on the local end and are uploaded to the global side to calculate the similarity between the new user and the candidate users. When the loss function obtained by training the aggregated parameters on the global side decreases more compared to the original loss function, it means that the similarity between the candidate user and the new user is higher, and the training parameters of the user are considered to have greater guiding significance. Conversely, when the loss function decreases less or even increases, the similarity is considered to be lower, and the guiding significance of the training parameters of the user is smaller. From this, the m users with the highest similarity are selected to assist the new user in completing the throughput prediction task. Considering the real-time requirement of prediction, in the initial stage, the throughput prediction uses a basic multilayer perceptron model MLP. According to the similarity, the attention weights of the parameters of the candidate users are set, with higher weights for users with greater guiding significance. The prediction is completed through weighting.

[0067] S3: Volatility-residual feature extraction based on TDLinear network model

[0068] As the length of use increases, new user historical data gradually accumulates, and when it exceeds a certain threshold, it can be trained independently. At this time, the TDLinear model is designed to extract the features of the volatility part and the stable part of the throughput in the historical data, so as to more accurately complete the prediction. When the user is more active in using the cloud desktop, the throughput presents a sharp fluctuation, and when the cloud desktop is idle, the throughput is relatively stable. This model divides the sliding time window, measures the fluctuation degree of the throughput by calculating the variance in the window, and thus extracts the two parts of features respectively. The memory module and the EWC review historical information are used for the two parts of features respectively.

[0069] S4: Model update strategy based on adaptive memory module and EWC

[0070] Based on the extracted volatility part features, an adaptive memory module is designed to store old patterns and update new patterns. The input data is passed through a fully connected layer, so that the features are retained on a few neurons. In order to save the patterns under different windows, the memory module is used to store the feature neurons of different patterns, and the Pearson correlation coefficient is used to measure the correlation between the feature neurons in the current window and the neurons in the memory module. If the correlation is high, the memory module neurons are used to complete the training directly, and if the correlation is low, the memory module is stored to complete the update. For the foregoing stable part features, the model parameters of the sliding window are constrained, and the same penalty term is used to review the previous training parameters when the constraint is used, so as to achieve the purpose of constantly reviewing old knowledge.

[0071] S5: Throughput prediction solution for user long-term use stage training forgetfulness

[0072] After the TDLinear model feature is divided, the fluctuation feature F act and the stable feature F idle when idle are extracted respectively. The fluctuation feature F act uses the memory module to realize adaptive update by constantly measuring the correlation of neurons, uses Dropout regularization to improve the generalization ability of the network, prevents model overfitting, uses Ensemble to balance the prediction results of different memory units, and finally uses a linear prediction model to obtain the prediction result y act of the fluctuation feature part. For the stable feature F idle , the EWC reviews the historical training parameters, in order to retain the knowledge of the previous task, the key parameters are punished, so that they will not easily change when training a new task. The same linear model is used to obtain the prediction result y idle . When outputting, the two parts are aggregated in the time dimension to obtain the final output result wherein y act is a first prediction result; y idle is a second prediction result; y pred is a final output result.

[0073] From the above description, it can be seen that the above-mentioned embodiments of the present application achieve the following technical effects:

[0074] Through the federated learning algorithm, the existing throughput history use knowledge of other users is used to assist training under the premise of ensuring user data privacy. With long-time use, when the data volume accumulates enough, the user will independently use the historical data to complete the training. The present scheme continuously reviews the old mode based on continuous incremental prediction, and designs a memory module to incrementally update the new mode, which has the technical effect of avoiding the occurrence of forgetting in training. The two stages are combined to realize the whole life cycle throughput prediction of the cloud desktop user.

[0075] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A cloud desktop throughput prediction method for cold start and pattern forgetting, characterized by: include: Step 1: Similarity-based attention federated learning assisted training; Step 2: Throughput prediction for the initial secure cold start of new users; Step 3: Fluctuation-residual feature extraction based on TDLinear network model; Step 4: Model update based on adaptive memory module and EWC; Step 5: Throughput prediction for training and forgetting during the long-term user usage phase; In step 1, first obtain the prior knowledge of the new user, which is generated when the user registers or binds an account, including basic identity information and usage information; In step 1, by comparing the user's prior knowledge, the new user is selected with higher relevance. The user groups serve as federated candidate sets. During this process, users encrypt and upload their data to the global side, which performs the screening operation. The same lightweight prediction model, multi-layer perceptron, is used on the global side and the local user side. The federated candidate sets are trained separately and the gradient parameters during the training process are encrypted and uploaded to the global side. At the same time, new users are also trained using a small amount of existing data, and the training parameters are also encrypted and uploaded. The global side calculates the similarity of the training parameters obtained in step 1, and obtains the similarity weight through two rounds of interaction with the Federal Reserve users. The updated training parameters are calculated based on the weights and returned to the new user to update the model. During this process, the new user only interacts with the global side once and needs to wait for the federated candidate set to interact with the global side. After obtaining the training loss before and after the two training updates on the local side in step 2, upload it to the global side to calculate the similarity between the new user and the candidate user. When the loss function obtained by the parameter training after aggregation on the global side is reduced more than the original loss function, it means that the similarity between the candidate user and the new user is higher, and it is considered that the training parameters of the user have greater guiding significance. On the contrary, when the loss function is reduced less or even increased, it is considered that the similarity is lower, and the training parameters of the user have less guiding significance. The user with the highest similarity is selected from them. The user is used to assist new users in completing the throughput prediction task. Considering the real-time requirements of the prediction, the throughput prediction uses a basic multi-layer perceptron model in the initial stage. The attention weights are set for the parameters of candidate users based on similarity. The weights with greater guiding significance are higher. The prediction is completed in a weighted manner. In step 3, as the usage time increases, new user historical data gradually accumulates. When it exceeds a certain threshold, independent training can be completed. At this time, a TDLinear model is designed to extract the features of the fluctuating and stable parts of the throughput in the historical data. When the user is more actively using the cloud desktop, the throughput fluctuates violently, while the throughput is relatively stable when the cloud desktop is idle. The sliding time window is divided and the variance within the window is calculated to measure the degree of throughput fluctuation, thereby extracting two parts of features respectively. The memory module and EWC are used to review historical information for the two parts of features respectively. In step 4, based on the extracted fluctuation part features, the old mode is stored and the new mode is updated through the adaptive memory module, the input data is passed through the fully connected layer, so that the features are retained on a small number of neurons, the memory module is used to store feature neurons of different modes, and the Pearson correlation coefficient is used to measure the correlation between the feature neurons under the current window and the existing neurons in the memory module. If the correlation is high, the memory module neurons are directly used to complete the training. If the correlation is low, the memory module is stored to complete the update. For the stable part features, the model parameters of the updated sliding window are constrained, and the same penalty term is used to review the previous training parameters and continuously review the old knowledge during the constraint; In step 5, after the division of the TDLinear model features, the fluctuation features representing the user's active use of the cloud desktop and the stable features when idle are extracted separately. The fluctuation features are adaptively updated by continuously measuring the correlation of neurons using the memory module. Dropout regularization is used to improve the generalization ability of the network and prevent model overfitting. Ensemble is used to balance the prediction results of different memory units. Finally, the linear prediction model is used to obtain the first prediction result of the fluctuation feature part. ; In step 5, for stable features, EWC is used to review historical training parameters. In order to retain the knowledge of previous tasks, a penalty is imposed on key parameters so that they will not change easily when training new tasks. The same linear model is used to obtain the second prediction result. , when outputting, aggregate the two parts in the time dimension to obtain the final output result ;in, is the first prediction result; is the second prediction result; Final output result.

2. The cold start and pattern forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 1, an overall framework of horizontal federated learning is adopted, which includes a global side and a local user side. The global side is used to process the data uploaded by each user, and the results obtained after joint processing are returned to each local user.

Images