Cloud desktop throughput prediction method for cold boot and mode forgetting

Abstract

The invention discloses a cloud desktop throughput prediction method for cold start and mode forgetting. The method comprises the steps of 1, learning aided training based on similarity attention federal; 2, throughput prediction for safe cold start in the initial stage of a new user is carried out; step 3, fluctuation-residual feature extraction based on a TDLinear network model is carried out; 4, updating the model based on the adaptive memory module and the EWC; and step 5, training forgotten throughput prediction in the long-term use stage of the user. Through a federated learning algorithm, on the premise of ensuring user data privacy, existing throughput historical use knowledge of other users is utilized to assist training. An old mode is continuously reviewed based on continuous incremental prediction, a memory module is designed to incrementally update a new mode, and the method has the advantages of avoiding forgetting in training and integrating two stages to realize full-life-cycle throughput prediction of the cloud desktop user.

Description

Technical Field

[0001] The present invention relates to the technical field of video stream throughput prediction in a cloud desktop scenario, and in particular to a cloud desktop throughput prediction method oriented to cold start and mode forgetting. Background Art

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

[0003] Existing mature cloud desktop products and research mainly consider how to optimize encoding and decoding to improve user quality of use, and rarely pay attention to network issues related to throughput and bandwidth resource allocation. Network bandwidth resources are the key factors that determine the fluency and stability of cloud desktop services. Fixed bandwidth is difficult to meet the multi-scenario needs of users, which inevitably causes resource waste in an undersaturated state and limited network transmission in a saturated state. A real-time throughput prediction solution helps guide the dynamic allocation of bandwidth. Cloud desktop bandwidth prediction based on historical data mainly faces two problems: "cold start" and "training forgetting". When a new user uses the service for the first time, there is less historical data, which makes it difficult to support model training to achieve accurate throughput prediction. There is a "cold start" problem. A feasible solution is to combine existing knowledge or recommendation strategies. However, the difficulty in the cloud desktop scenario is: while utilizing the knowledge of other users, it is necessary to ensure data isolation between each other and achieve "secure start" under the premise of ensuring data privacy.

[0004] In addition, as the user usage time increases, a large amount of historical data accumulates. New knowledge will cover part of the old knowledge during long-term training, leading to the problem of "training forgetting". Constantly reviewing old knowledge is a feasible solution. During the user's use of cloud desktops, there will be frequent changes in usage scenarios. Another major problem is how to effectively avoid the new scenario knowledge learned during training from covering the old scenario and effectively utilize the time series features.

[0005] In summary, there is at least one of the following technical problems:

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

[0007] As users use the device for longer periods of time, a large amount of historical data accumulates. During long-term training, new knowledge will cover part of the old knowledge, leading to the problem of "training forgetting".

[0008] How to effectively avoid the new scene knowledge learned during training from covering the old scenes and effectively utilize the time series features. Summary of the invention

[0009] The main purpose of the present invention is to provide a cloud desktop throughput prediction method for cold start and pattern forgetting, so as to solve the two problems of "cold start" and "training forgetting" faced by cloud desktop bandwidth prediction based on historical data in the prior art. As the user usage time increases, a large amount of historical data accumulates. In the long-term training process, new knowledge will cover part of the old knowledge, leading to the "training forgetting" problem. How to effectively avoid the new scene knowledge learned during the training process from covering the old scene, and effectively utilize the technical problem of time series features.

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

[0011] Step 1: Similarity-based attention federated learning assisted training;

[0012] Step 2: Throughput prediction for the initial secure cold start of new users;

[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: Train the forgotten throughput prediction for the long-term user usage phase.

[0016] Preferably, 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.

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

[0018] Preferably, in step 1, by comparing the user's prior knowledge, m user groups with high correlation are selected for the new user as the federal candidate set. In this process, the users encrypt and upload their respective data to the global side, and the global side completes the screening operation. The same lightweight prediction model multi-layer perceptron is used on the global side and the local user side. The federal candidate set will be trained separately and the gradient parameters in the training process will be encrypted and uploaded to the global side. At the same time, the new user is also trained using a small amount of existing data, and the training parameters are also encrypted and uploaded.

[0019] Preferably, 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 user, obtains the updated training parameters based on the weight calculation, and returns them 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, after obtaining the training losses before and after two training updates at the local end in step 2, upload them to the global side to calculate the similarity between the new user and the candidate user. The more the loss function obtained by the parameter training after aggregation on the global side is reduced compared with the original loss function, the higher the similarity between the candidate user and the new user, and it is believed that the training parameters of the user have greater guiding significance. On the contrary, when the loss function is reduced less or even increases, it is believed that the similarity is lower, and the guiding significance of the training parameters of the user is smaller. The m users with the highest similarity are selected to assist the new user 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, and sets attention weights for the parameters of the candidate users according to the similarity. The weights with greater guiding significance are higher, and the prediction is completed in a weighted manner.

[0021] Preferably, in step 3, as the usage time increases, historical data of new users gradually accumulates. When it exceeds a certain threshold, the training can be completed independently. At this time, the TDLinear model is designed to extract the features of the fluctuating part and the stable part 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 degree of throughput fluctuation is measured by calculating the variance within the window, thereby extracting two parts of features respectively, and using the memory module and EWC to review historical information for the two parts of features respectively.

[0022] Preferably, in step 4, based on the extracted fluctuating 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 few 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 in 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. When 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 constraints.

[0023] Preferably, 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 respectively, the fluctuation features are adaptively updated by continuously measuring the neuron correlation using a memory module, the network generalization ability is improved by Dropout regularization to prevent the model from overfitting, Ensemble is used to balance the prediction results of different memory units, and finally the linear prediction model is used to obtain the first prediction result y of the fluctuation feature part act .

[0024] Preferably, 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 y idle , when outputting, aggregate the two parts in the time dimension to obtain the final output result Among them, y act is the first prediction result; y idle is the second prediction result; y pred Final output result.

[0025] The application of the technical solution of the present invention has the following technical effects:

[0026] Through the federated learning algorithm, while ensuring the privacy of user data, the existing throughput history usage knowledge of other users is used to assist training. With long-term use, when the amount of data accumulates enough, users will independently use their own historical data to complete training. This solution proposes to continuously review the old model based on continuous incremental prediction, and design a memory module to incrementally update the new model, which has the technical effect of avoiding forgetting during training. It integrates two stages to achieve the full life cycle throughput prediction of cloud desktop users. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 The overall flow chart of the cold start and mode forgetting oriented cloud desktop throughput prediction method according to the present invention is shown;

[0029] Figure 2 Shows Figure 1 The similarity-attention-based federated learning framework diagram used in the cloud desktop throughput prediction method for cold start and pattern forgetting;

[0030] Figure 3 Shows Figure 1 The elastic weight-preserving incremental learning model diagram adopted in the cloud desktop throughput prediction method for cold start and pattern forgetting;

[0031] Figure 4 Shows Figure 1 The fluctuation-residual feature extraction model diagram based on TDLinear is adopted in the cloud desktop throughput prediction method for cold start and pattern forgetting;

[0032] Figure 5 Shows Figure 1 Multi-layer perceptron structure model diagram of the cloud desktop throughput prediction method for cold start and pattern forgetting. DETAILED DESCRIPTION

[0033] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0034] like Figures 1 to 5 As shown, an embodiment of the present invention provides a cloud desktop throughput prediction method for cold start and pattern forgetting, including: step 1: similarity attention federated learning assisted training; step 2: throughput prediction for the initial safe cold start of new users; step 3: fluctuation-residual feature extraction based on the TDLinear network model; step 4: model update based on adaptive memory module and EWC; step 5: throughput prediction for training forgetting in the long-term use stage of users.

[0035] This embodiment comprehensively considers the entire life cycle of users using the cloud desktop. When the amount of new user data is insufficient and difficult to predict independently at the initial stage of use, this embodiment designs a federated learning algorithm to use the existing throughput history usage knowledge of other users to assist in training while ensuring user data privacy. With long-term use, when the amount of data accumulates sufficiently, users will independently use their own historical data to complete training. This solution proposes to continuously review old patterns based on continuous incremental predictions, and design memory modules to incrementally update new patterns to avoid forgetting during training. The two stages are combined to achieve throughput prediction for cloud desktop users throughout their life cycle. The two-stage prediction details of this solution are introduced in detail below.

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

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

[0038] In the cloud desktop scenario, each user uses the cloud desktop as an independent individual to complete their own business. They are confidential and unconnected to each other. When a new user starts using the cloud desktop, it is difficult to ensure the accuracy of throughput prediction by training with only a small amount of historical data. At this time, the historical usage of similar users is often instructive. In order to use the general rules of historical throughput usage of other users to assist training while ensuring the security of user data, this solution uses a federated learning framework to solve the "safe cold start" problem, and designs an attention mechanism based on user similarity to adjust the guidance weights of 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 the 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 the federal candidate set. In this process, users will encrypt and upload their own data to the global side, and the global side will complete the screening operation to protect their privacy. In order to ensure the speed of model training and the real-time prediction, the same lightweight prediction model multi-layer perceptron is used on the global side and the local user side. The federal candidate set U will be trained separately and the gradient and other parameters in the training process will be encrypted and uploaded to the global side. At the same time, new users will also use a small amount of existing data for training, and the training parameters will also be encrypted and uploaded. The global side will calculate the similarity of the obtained training parameters, and through two rounds of interaction with the federal reserve user U, obtain the similarity weight Similarity={s1,s2,…,s m}, based on the weight calculation, the updated training parameters are obtained 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 federated candidate set U to interact with 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 federated candidate set U to guide the new user user new Completing parameter updating and training is a key part. The present invention defines similarity based on the influence of different parameters on the gradient descent of the training loss function. Through two rounds of iterations, the similarity weight set between 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 new user user new , define its training model M new After training is completed, the parameters θ are generated new and loss L new , the training model set of the Federal Reserve user U The set of parameters generated by training and loss Parameter θ i After the global model M global The parameters after aggregation are First, the global side aggregates the parameters in the model:

[0043]

[0044] in, is the parameter updated after training; M global is the global model; θ new Parameters generated after training for new users; A collection of training models for the Federal Reserve user U The set of parameters generated by training.

[0045] Here, the parameters of the new user and the parameters in the Federal Reserve are aggregated and trained separately to obtain the updated parameters after training. These parameters are transferred back to the local machine for secondary training, and the loss updated after training is Pass it to the global side again and calculate the similarity:

[0046]

[0047] Among them, s i Represents the similarity of user i to the new user in the federated candidate set. Similarity set Similarity = {s1, s2, …, s m}; For normalization operation, the result is normalized to the range of [-1,1]; L i and L i 'represent the original and updated losses of user i, M i and M global Represent the training model and global model of the i-th user, θ i and θ new Respectively represent the training parameters of user i and the parameters generated after the new user is trained; B is the training batch size. At this time, the training parameters of the new user after updating according to the similarity attention are:

[0048]

[0049] in, represents the training parameters updated using similarity attention, λ is the weight of the parameters of the new user after aggregation, and θ new The parameters generated for training new users, s i is the similarity between user i and the new user in the federated candidate set, is the training parameter of user i on the server side; in the proposed similarity calculation scheme, when the loss function obtained by parameter training after global aggregation is reduced more than the original one, it means that the similarity between the user and the new user is higher, and this scheme will consider that the training parameters of the user have greater guiding significance. On the contrary, when the loss is reduced less or even increased, it is considered that the training parameters of the user have less guiding significance, and the weight is set to 0. Through the above iterative process, the parameter update of the new user is guided by combining the federal candidate user set U.

[0050] 2. Phase 2: Training forgetting solution based on continuous incremental learning

[0051] After a user has used the system for a certain period of time, his / her historical bandwidth usage record will increase. When the accumulated data exceeds the threshold value τ, the system will no longer combine the data of other users, but use its own historical throughput data to complete the prediction through independent training. Considering that the new mode will have a certain coverage impact on the old mode during training, this solution proposes a continuous incremental learning model to complete training and prediction. The input throughput data is divided into a fluctuating part and a remaining part. The fluctuating part belongs to the part where the user is actively using the cloud desktop, and the remaining part is considered to be the idle state of the user. The features of the two parts are extracted separately for prediction, and finally they are spliced ​​according to time to obtain the prediction result. The whole process is updated using an adaptive incremental model.

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

[0053] First, the feature extraction rules of the scheme for the fluctuating part and the remaining part of the input data will be introduced. The preprocessing of the input data is performed in a sliding window manner. Before training, the data is divided into the fluctuating part and the remaining part according to the user's cloud desktop usage status. Through data cleaning, it is found that when the user uses the cloud desktop to complete related tasks, the bandwidth will fluctuate more violently, while 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 user's throughput fluctuation patterns and general stable patterns, that is, the throughput characteristics of all services when the user uses the cloud desktop are obtained. Specifically, this 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's active use of the cloud desktop, and the part less than is considered to be the general rule of the user's use of the service. After dividing the input data into two parts, the data is respectively extracted through the fully connected layer. The number of input neurons is reduced through the fully connected layer, and the features are retained on a few neurons as much as possible. In order to save the patterns under different windows, this scheme designs a memory module to store the feature neurons of different patterns, and uses the Pearson correlation coefficient to measure the similarity between the feature neurons under the current window and the neurons in the memory module:

[0054]

[0055] Among them, r xy is the Pearson coefficient, n is the number of neurons, x i is the i-th neuron in the current window, y iis the i-th neuron in the memory module. When the feature neuron in the sliding window is similar to the existing one in the memory module, the training of the sliding window will be skipped directly, the sliding window will be moved forward, and the memory module will be used for training directly. When the feature neuron has never appeared in the memory module, it will be stored in the memory module and further embedded; for the remaining part (stable part), elastic weight consolidation (EWC) is used to complete further operations.

[0056] 2.2 Incremental Continuous Learning Based on EWC

[0057] For fluctuating throughput data, when the data in the sliding window is considered to have a new pattern through the memory module, it is necessary to use an embedding method to aggregate the feature vector of the new pattern with the original feature vector to complete the learning process of the memory module. The present invention stores the data in the sliding window as new data in the memory module, and at the same time, the new feature vector is Dropouted in the same proportion and then spliced ​​with the original feature vector to ensure that the feature vector ratio retained in each window is the same. For the remaining part of the data, a similar processing method as the fluctuating feature is adopted. The difference is that, considering that the training difference of the remaining part should not be too large, the present invention constrains the model parameters of the updated sliding window each time, and uses the EWC strategy when constraining. 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 weights to each parameter according to the importance; at the same time, constraints are imposed on the parameters when learning a new task, so that it will not change too much, thereby achieving the purpose of retaining 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 to measure the uncertainty of parameter estimation. At this time, the parameter θ i The importance weight can be calculated by the following formula:

[0058]

[0059] Among them, F i,t is the parameter θ i Fisher information matrix for task t. EWC also constrains the elastic weights of new task parameters through regularization terms:

[0060]

[0061] Among them, the new task is B and the old task is A. B (θ) represents the loss function of directly training the new task without adding EWC, F i represents the Fisher information matrix, θ irepresents the network parameters when training a new task B, θ A,i * represents the optimal parameters obtained by completing the old task A, and λ is the regularization coefficient. By adding the EWC constraint to the loss function, the model can protect the important parameters learned previously when learning new tasks, avoiding excessive adjustment of parameters to cover old knowledge.

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

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

[0064] This strategy mainly consists of the local user side and the global side. During training, the local user only uploads the model training parameters to the global side, ensuring data security and privacy. new For example, when registering for a cloud desktop service, some basic information is often included. prior , by comparing this information, a federated user candidate set U with similar information to the new user is determined. On the local side, the new user and the candidate user U first adopt a common network structure and use their respective data for the first round of training, generating a training loss and training parameters The model training parameters are uploaded to the global side. On the global side, the parameters of U are aggregated with the new users one by one, and the second round of training is completed together. At the same time, the model parameters are updated and transmitted back to the local side. The local side uses the updated parameters to complete the third round of training, generating updated losses. The influence of user parameters in the candidate set is measured by comparing the changes in the loss function.

[0065] S2: Throughput prediction solution for safe cold start in the initial use phase for new users

[0066] Get the training loss before and after two training updates on the local side and After that, it is uploaded 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 compared with the original loss function, it means that the similarity between the candidate user and the new user is higher, and it is believed 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 believed that the similarity is lower, and the guiding significance of the training parameters of the user is smaller. The m users with the highest similarity are selected to assist the new user in completing the throughput prediction task. Considering the real-time requirements of the prediction, the basic multi-layer perceptron model MLP is used for the throughput prediction in the initial stage. Attention weights are set for the parameters of the candidate users according to the similarity. The weights with greater guiding significance are higher, and the prediction is completed in a weighted manner.

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

[0068] As the usage time increases, the historical data of new users gradually accumulates, and when it exceeds a certain threshold, the training can be completed independently. At this time, the TDLinear model is designed to extract the features of the fluctuating and stable parts of the throughput in the historical data, so as to facilitate more accurate prediction. When users are more actively using the cloud desktop, the throughput fluctuates violently, while the throughput is relatively stable when the cloud desktop is idle. This model divides the sliding time window and measures the degree of throughput fluctuation by calculating the variance within the window, thereby extracting two parts of features respectively. The memory module and EWC are used to review historical information for the two parts of features.

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

[0070] Based on the aforementioned extracted fluctuation 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 small number of neurons. In order to preserve the patterns under different windows, the memory module is used to store the feature neurons of different patterns. 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, they are stored in the memory module to complete the update. For the aforementioned stable features, the model parameters of the updated sliding window are constrained, and the same penalty term is used to review the previous training parameters during the constraint, so as to achieve the purpose of constantly reviewing old knowledge.

[0071] S5: Throughput prediction solution for training forgetfulness during long-term user usage

[0072] After the TDLinear model features are divided, the fluctuation characteristics F representing the active use of cloud desktop by users act and the stable characteristic F at idle idle are extracted separately. Fluctuation feature F act The memory module is used to achieve adaptive updates by continuously measuring the correlation of neurons. Dropout regularization is used to improve the generalization ability of the network and prevent overfitting of the model. Ensemble is used to balance the prediction results of different memory units. Finally, the linear prediction model is used to obtain the prediction result y of the fluctuation feature part. act For stable feature F idle , 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 prediction result y idle When outputting, the two parts are aggregated in the time dimension to obtain the final output result. Among them, y act is the first prediction result; y idle is the second prediction result; y pred Final output result.

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

[0074] Through the federated learning algorithm, while ensuring the privacy of user data, the existing throughput history usage knowledge of other users is used to assist training. With long-term use, when the amount of data accumulates enough, users will independently use their own historical data to complete training. This solution proposes to continuously review the old model based on continuous incremental prediction, and design a memory module to incrementally update the new model, which has the technical effect of avoiding forgetting during training. It integrates two stages to achieve the full life cycle throughput prediction of cloud desktop users.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cloud desktop throughput prediction method for cold start and mode forgetting, characterized in that: 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: Train the forgotten throughput prediction for the long-term user usage phase.

2. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In the 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.

3. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 1, firstly, the prior knowledge of the new user is obtained, and the prior knowledge is generated when the user registers or binds an account, including basic identity information and usage information.

4. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 1, by comparing the user's prior knowledge, m user groups with high correlation are selected for the new user as the federal candidate set. In this process, the users encrypt and upload their data to the global side, and the global side completes the screening operation. The same lightweight prediction model multi-layer perceptron is used on the global side and the local user side. The federal candidate set will be trained separately and the gradient parameters in the training process will be encrypted and uploaded to the global side. At the same time, the new user is also trained using a small amount of existing data, and the training parameters are also encrypted and uploaded.

5. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 1, the global side calculates the similarity of the obtained training parameters, and obtains the similarity weight through two rounds of interaction with the federal reserve user. 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.

6. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: After the training losses before and after two training updates are obtained on the local side in the step 2, they are uploaded to the global side to calculate the similarity between the new user and the candidate user. The more the loss function obtained by the parameter training after aggregation on the global side is reduced compared with the original loss function, the higher the similarity between the candidate user and the new user, and it is believed that the training parameters of the user have greater guiding significance. On the contrary, when the loss function is reduced less or even increases, it is believed that the similarity is lower, and the guiding significance of the training parameters of the user is smaller. The m users with the highest similarity are selected to assist the new user in completing the throughput prediction task. Considering the real-time requirements of the prediction, the basic multi-layer perceptron model is used for the throughput prediction in the initial stage. Attention weights are set for the parameters of the candidate users according to the similarity. The weights with greater guiding significance are higher, and the prediction is completed in a weighted manner.

7. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 3, as the usage time increases, historical data of new users gradually accumulates. When it exceeds a certain threshold, the training can be completed independently. At this time, the TDLinear model is designed to extract the features of the fluctuating part and the stable part 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 degree of throughput fluctuation is measured by calculating the variance within the window, so as to extract two parts of features respectively. The memory module and EWC are used to review the historical information for the two parts of features respectively.

8. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 4, based on the extracted fluctuating 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 few 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 in 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. When 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 constraints.

9. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: 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 respectively. The fluctuation features use the memory module to achieve adaptive update by continuously measuring the correlation of neurons, and the Dropout regularization is used to improve the generalization ability of the network to prevent the model from overfitting. Ensemble is used to balance the prediction results of different memory units, and finally the linear prediction model is used to obtain the first prediction result y of the fluctuation feature part. act .

10. The cold start and mode forgetting cloud desktop throughput prediction method according to claim 1, characterized in that: In step 5, for stable features, EWC is used to review historical training parameters. In order to retain the knowledge of previous tasks, penalties are 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 y idle , when outputting, aggregate the two parts in the time dimension to obtain the final output result Among them, y act is the first prediction result; y idle is the second prediction result; y pred Final output result.

Images