An edge caching method based on content popularity

Abstract

The present invention seeks to protect an edge caching method based on content popularity, comprising the following main steps: S1, constructing a relevant system model based on server clusters, devices and user request content; S2, classifying historical data of different lengths and using corresponding algorithms to predict the popularity of content at future times; S3, calculating the total revenue value of each content based on the popularity obtained in S2 and the size of the content; S4, for each server cluster, calculating the content sequence with the highest benefit under the current capacity; S5, obtaining the result of whether each content is cached in the edge server cluster. Compared with existing methods, the method of the present invention has the following main advantages: (1) Different algorithms are used for prediction of historical data of different lengths, so that the error of the prediction result is smaller. (2) The content cache placement problem is converted into a benefit value maximization problem, and content with larger total benefits can be cached in the edge server cluster.

Description

Technical Field

[0001] The present invention relates to the field of mobile edge computing, and in particular to an edge content caching method based on content popularity. Background Art

[0002] Now, the fifth generation of mobile communications has attracted much attention worldwide and has been officially commercialized. The core technologies of 5G mainly include millimeter wave and massive multi-input multi-output. In the future development, increasing users' communication mode choices and further improving users' network experience are important goals of 5G evolution. The emergence of 5G means that users' demand for video streaming services will further increase, which will bring huge challenges to the streaming service system. The collaborative-edge cloud computing architecture that combines cloud computing and edge computing can effectively reduce the burden of the streaming service system. Streaming content caching under the collaborative edge cloud computing architecture is of great significance to improving user service quality. Caching technology can reduce latency, increase transmission rate and save backhaul link consumption. It is one of the important technologies to cope with the exponential growth of mobile data traffic and will inevitably play a key role in the future 5G system.

[0003] In traditional networks, the content requested by mobile users is usually served by remote Internet content servers provided by content providers. When users retrieve the same popular content from remote servers, the remote servers need to send the same files repeatedly, which may result in a large amount of repeated traffic and network congestion during peak network periods. In mobile edge cache networks, by using the storage services provided by mobile edge computing servers, popular content can be cached closer to users, which can reduce repeated content transmission from content servers to edge cache nodes, effectively alleviate network congestion problems, and effectively reduce content retrieval delays and transmission energy consumption. However, in current edge computing caching methods, the factors considered are usually very limited. The caching method is closely related to factors such as the popularity of the content to be cached and user mobility. Therefore, studying the selection method of cached content in edge computing is crucial to reducing transmission delays and improving end-user experience. In addition, the global content popularity is different from the content of interest to the end users served by each micro base station, so it is also necessary to study the placement method of cached content so that each base station can cache as much of the content most needed by users as possible within a limited cache space.

[0004] Service providers can deploy streaming content on edge servers. Edge servers are responsible for responding to users' content requests, reducing transmission delays, and ensuring user service quality. At the same time, streaming service providers can also dynamically use the powerful computing resources of cloud servers according to actual needs. Therefore, it is necessary to design corresponding caching methods based on the frequency of content requests from end users to reduce service provider costs and user content acquisition delays. However, the popularity of content access changes at any time. Therefore, it is necessary to find a content caching method that takes content popularity into account so that edge service providers can cache content at a lower cost while improving the hit rate of content and meeting users' needs for low latency in acquiring content.

[0005] Although the storage capacity of edge servers is much higher than that of terminal devices, it is ultimately limited. Therefore, it is necessary to cache content that is of great benefit to edge servers as much as possible. First, consider the popularity feature. After establishing a relevant model for the user's behavioral characteristics, future behavior can be predicted based on specific historical data. At the same time, the impact of content size is taken into account. For example, larger content will cause other content to be unable to be cached and take longer to obtain. At the same time, set the popularity gain and delay gain, consider the proportion of these two factors, and finally get the total value benefit of each content. In the subsequent cache placement plan, the goal is to select content with large total value benefits for caching. This is an NP-hard problem. The problem is converted into a benefit maximization problem. Under the condition of meeting the constraints, the most valuable batch of content can be cached, which ultimately reduces the delay overhead of users obtaining content and improves the hit rate.

[0006] The application publication number is CN113965937A, a method for predicting content popularity based on clustering federated learning in a fog wireless access network, characterized in that the prediction method comprises the following steps: step S1, according to the neighborhood set of local users and content, using the local user information and content information obtained by the fog access point, constructing the initial features of each local user and each content through information fusion; step S2, taking the initial features of the local user and content as input, and then taking the content request probability of the local user as the prediction target, establishing a prediction model based on a dual-channel neural network for each fog access point, and setting the binary cross entropy loss as the loss function to optimize the model parameters; step S3, using the clustering federated learning method, performing distributed training on the prediction model of each fog access point, and adaptively training the fog access points with similar area types. Clustering is performed to specialize the model parameters for each fog access point; step S4, using the number of historical requests to obtain the activity of local users, and obtain the prediction result of local popularity based on the activity of local users and the predicted content request probability; step S5, based on the content information, taking the content request probability of mobile users as the prediction target, establish the objective function of the preference model for each mobile user, and set the optimization target of the preference model to minimize the objective function; step S6, each fog access point collects the content request probability of the mobile users currently associated with it, and obtains the prediction result of mobile popularity; step S7, each fog access point normalizes the local popularity and mobile popularity of all content respectively, and integrates the local popularity and mobile popularity according to the current proportion of mobile users, to obtain the content popularity of each fog access point.

[0007] The above invention establishes features by collecting user information and content information, predicts the popularity of content by establishing a dual-channel neural network model, and then optimizes the model parameters. It does not consider that the use of neural networks for prediction of data with shorter historical lengths will cause too much time overhead. If some new content enters the network, there is only a small amount of historical data, which has not reached the number of input data of the neural network, which will cause the prediction accuracy to decrease. At the same time, the above invention takes the content request probability of mobile users as the prediction target, and establishes a preference model objective function for each mobile user. However, the influence of the behavior of a single user on the group is minimal. The local group behavior should be considered during the prediction, and the choice of most people can represent the popularity of the content in the current area. The present invention takes the above two points into consideration, selects the behavior of most users in a server cluster, obtains the popularity of the content after collection, and classifies the content of different lengths. For content with less historical data, an improved exponential smoothing algorithm is used for prediction, and a high-accuracy result can be obtained in a shorter time. For the content popularity with sufficient historical data, a dynamic optimization LSTM neural network is used for prediction, and the parameters are continuously optimized so that the deviation between the obtained result and the true value is smaller. Summary of the invention

[0008] The present invention aims to solve the problems of content popularity prediction and edge server caching methods in existing edge computing, and proposes a classification prediction algorithm to reduce the content popularity prediction error. At the same time, a related algorithm is used to allow the edge server to cache the content with the greatest benefit in a limited storage space, so that the cache hit rate of users obtaining content is improved and the delay of obtaining content is reduced. A method of edge caching based on content popularity is proposed. The technical solution of the present invention is as follows:

[0009] An edge caching method based on content popularity comprises the following steps:

[0010] S1, build an end-edge cloud system model based on server clusters, devices and user request content, obtain the number of requests for each content in different time periods, and regard the number of requests per unit time by terminal users as the historical popularity of the content in the unit time period;

[0011] S2, classify historical data of different lengths, and use the improved exponential smoothing prediction algorithm and dynamic optimization long short-term memory algorithm to predict the popularity at future moments after classification;

[0012] S3, the popularity predicted in step S2 is regarded as popularity gain. Since the delay of obtaining content is related to the size of the content itself, if the content is cached in the edge server cluster, the reduced delay compared to obtaining it from the cloud is regarded as the delay cost gain, and the total value benefit of each content is calculated;

[0013] S4: For each server cluster, a cache placement algorithm is used to calculate the most efficient content sequence under the current capacity, and the content corresponding to this batch of content sequences is cached in the server;

[0014] S5, obtaining the result of whether each content is cached in the edge server cluster.

[0015] Furthermore, in step S1, the method for constructing the end-edge-cloud system model and obtaining the historical popularity specifically includes:

[0016] Most content is stored in the cloud, and the edge server cluster selects some content for storage. The terminal user initiates a request and divides the environment into three types: cloud, edge, and terminal. The cloud server is a device that is far away from the user terminal. Each edge server cluster serves some terminal devices. The terminal devices will initiate requests for certain content at different time periods. The edge server collects historical data on the requested content and obtains the content popularity at different time periods, thereby obtaining a historical popularity set of all content under the edge server cluster for subsequent popularity prediction.

[0017] Further, in step S2, historical data of different lengths are classified. After classification, an improved exponential smoothing prediction algorithm and a dynamic optimization long short-term memory algorithm are used to predict the popularity to predict the popularity at a future moment, specifically including:

[0018] Set lp as the length of the content set and set the length limit LB;

[0019] (1) If lp < LB, it is classified as new content and predicted using an improved exponential smoothing prediction algorithm;

[0020] (2) If lp ≥ LB, the dynamic optimization long short-term memory algorithm is used to predict the popularity.

[0021] Further, in step (1), an improved exponential smoothing prediction algorithm is used for prediction. In the traditional exponential smoothing prediction method, the weight factor α determines the size of the final predicted value. When α is larger, the popularity closer to the current time period has a greater impact on the prediction result. Conversely, when the value of α is smaller, the popularity of older content has a greater impact. The formula is as follows.

[0022]

[0023] represents the content popularity predicted at time t using the exponential smoothing prediction algorithm, where n at the moment, and represents the content popularity at the previous moment, and subsequent operations need to be carried out based on the popularity at the previous moment. t n represents the current time period, and t s represents the starting time period. α is a variable factor that is fixed and ranges from [0, 1], and is used to determine the influence degree of the popularity in the previous time period on the future time period. Since the weight factor α in the traditional method is static, it makes the smoothing prediction difficult to conform to the changes of the time series itself. Therefore, the present invention uses an exponential smoothing prediction method with a dynamically changing weight factor, and its formula is as follows.

[0024]

[0025] represents the content popularity predicted at time t using the improved exponential smoothing prediction algorithm. Compared with the traditional method, in the improved method, the value of the weight factor α will change dynamically with n the change of t n and is no longer a fixed value. It is dynamically corrected over time, making the prediction accuracy closer to the true value.

[0026] Further, in step (2), using the dynamic optimization long short-term memory algorithm to predict the popularity specifically includes:

[0027] For content with long historical data, i.e. old content, dynamically optimized LSTM is used for prediction, and the time window size, dense layer number, LSTM network layer number and neuron number are used as optimization objects. m time window candidate values, k dense layer candidate values, n LSTM layer candidate values ​​and u neuron candidate values ​​are set, and the average percentage error MAPE of the LSTM model verification data obtained after the training times reach the limit is selected to measure the quality of the parameters; the definition of MAPE is as follows:

[0028]

[0029] y′ n is the value predicted by the model, y n is the true value of the sample, N is the number of samples, and M(m,k,n,u) represents the average percentage error obtained after prediction by the LSTM model constructed with the current parameters. Through continuous iterative training, a set of parameters with the best performance of M(m,k,n,u) in the test set is selected, and then the model used for old content prediction is determined; at the same time, the optimal m time window values ​​obtained are the standard LB for dividing new and old content.

[0030] Furthermore, in step S3, the popularity predicted in step S2 is regarded as popularity gain. Since the delay of obtaining content is related to the size of the content itself, if the content is cached in the edge server cluster, the reduced delay compared to obtaining from the cloud is regarded as the delay cost gain, and the total value benefit of each content is calculated; the specific formula is shown in the formula;

[0031] G i =A*ΔL i +B*P i

[0032] Among them, A and B are two variable factors A+B=1, which is used to control the proportion of the two types of income in the total value income; ΔL i Represents the latency gain of obtaining content, P i represents the popularity gain of the content, G i It represents the total value of the content.

[0033] Furthermore, in step S4, the cache placement problem is transformed into a benefit value maximization problem. The original problem makes a decision for each content, 0 represents no selection, 1 represents selection, which is a 0-1 variable, and it is a constrained 0-1 planning problem, which is an NP-hard problem. By considering the content that the cache can bring the greatest benefit to the server cluster, the problem is transformed into a benefit maximization problem, and the problem is transformed into the following formula:

[0034]

[0035] That is, the content with the largest total gain is selected from the collection for caching.

[0036] The advantages and beneficial effects of the present invention are as follows:

[0037] 1. Regarding popularity prediction related issues, existing related methods consider predicting the user's movement path to obtain the potential predictability of user mobility, thereby obtaining content that may be popular. There are also block-based video popularity prediction methods, but most of these methods do not consider the impact of historical data of different lengths on the prediction results, and do not consider reducing the prediction error. The present invention takes this into account when making predictions, and by establishing an end-edge cloud system, collects historical information of the content to define the property of content popularity, classifies historical data according to different lengths, and uses improved algorithms to predict the popularity of the content at future moments, effectively reducing the error between the predicted results and the true value.

[0038] 2. For historical data of different lengths, the present invention selects different methods for prediction in step S2. For historical data of shorter length, an improved exponential smoothing prediction algorithm is proposed. The weight factors in traditional exponential smoothing predictions are often determined and no longer changed, while the present invention proposes weight factors that change dynamically over time. The weight factor values ​​at different times are different, which is more in line with the real scene, so that the popularity predicted for short historical data is closer to the real popularity. For long historical data, by setting a long short-term memory model, the training data is input into the training to obtain a model with the lowest average percentage error, so that the time window can be selected as the input, which is also the dividing point of long and short historical data. For historical data of different lengths, a more accurate method can be used for prediction, and the subsequent server cluster can make more prepared choices based on this.

[0039] 3. Based on the above problems, a caching method based on content popularity is proposed. The predicted popularity is regarded as popularity gain, and the reduced acquisition delay after a certain content is cached in the edge server cluster is regarded as delay cost gain. The content cache placement problem is transformed into a benefit value maximization problem. The algorithm selects the content that maximizes the total benefit of the server for caching, and fully utilizes the storage resources of the server. At the same time, it improves the hit rate of users obtaining content and reduces the delay cost of obtaining content. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a cloud-edge system model diagram constructed by the preferred embodiment provided by the present invention;

[0041] Figure 2 It is a flow chart of an edge caching method based on content popularity of the present invention; DETAILED DESCRIPTION

[0042] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0043] The technical solution of the present invention to solve the above technical problems is:

[0044] The present invention provides an edge caching method based on content popularity, comprising the following steps:

[0045] S1, builds the edge-cloud system model based on the server cluster, devices, and user request content, such as Figure 1 As shown, let ES = {1, 2, 3, ..., m} represent the set of edge server clusters, CS = {c1, c2, c3, ..., cn}, represent the set of requested content, and TD = {1, 2, 3, ..., q} represent the set of terminal devices. The number of requests for each content in different time periods is obtained by counting and regarded as its popularity.

[0046] S2, classify historical data of different lengths into new content and old content according to the standard, and use the corresponding algorithm to predict the popularity of the content in the future;

[0047] S3, based on the popularity predicted in S2, regards it as popularity gain, regards the reduced latency after a certain content is cached in the edge server cluster compared to obtaining it from the cloud as latency gain, and calculates the total value benefit of each content;

[0048] S4, for each server cluster, use the cache placement algorithm to calculate the most efficient content sequence under the current capacity, obtain the server's cache decision for the content, and cache the corresponding content in the server;

[0049] S5: Based on the final result, the server cluster chooses to cache or not cache each content. When the content cached in the server cluster is requested by the user, it can be directly obtained from the edge server without sending a request to the cloud.

[0050] Furthermore, in step S1, the method for constructing the end-edge-cloud system model and obtaining the historical popularity specifically includes:

[0051] (1) According to relevant information, the cloud, edge, and terminal environments are divided. The cloud server is a device far from the user terminal. Each edge server cluster serves some terminal devices. The terminal devices will initiate requests for certain content at different time periods. The edge server collects the historical data of the request content. ES = {1, 2, 3, …, m} represents the set of edge server clusters, CS = {1, 2, 3, …, n} represents the set of request content, TD = {1, 2, 3, …, q} represents the set of terminal devices. Let t j = T j - T j-1 represent the unit time period. Let represent whether the device initiated a request for content c j ∈ CS during the time period t i . The popularity of content c i under server S during the time period t j can be expressed as:

[0052]

[0053] By this step, the popularity of content in different time periods is obtained, and thus the historical popularity set of all content under the edge server cluster is obtained for subsequent popularity prediction.

[0054] Furthermore, in step S2, classifying historical data of different lengths according to criteria and using corresponding algorithms for prediction includes:

[0055] (1) Use to represent the set of all historical data of content c ∈ CS in the edge server cluster s ∈ ES during the time period t1 to t n . Set lp as the length of the content set and set the length limit LB; if lp < LB, it is classified as new content and the improved exponential smoothing prediction algorithm IESP is used for prediction; if lp ≥ LB, the dynamic optimization long short-term memory algorithm DO-LSTM is used to predict the popularity.

[0056]

[0057] (2) For new content prediction, an improved exponential smoothing prediction algorithm is used for prediction. In the traditional exponential smoothing prediction method, the weight factor α determines the size of the final predicted value. When α is larger, the popularity closer to the current time period has a greater impact on the prediction result. Conversely, when α takes a smaller value, the popularity of older content has a greater impact. Its formula is shown as follows.

[0058]

[0059] represents the prediction obtained using the exponential smoothing prediction algorithm at time tn The popularity of the content at the moment, where Represents the popularity of the content at the last moment. Subsequent operations need to be performed based on the popularity at the last moment. n Represents the current time period, t s represents the starting time period, α is a variable factor, fixed and unchanged, with a value range of [0,1], which is used to determine the influence of the popularity of the previous time period on the future time period. Since the weight factor α in the traditional method is static, it is difficult for the smooth prediction to conform to the changes of the time series itself. Therefore, the present invention uses an exponential smoothing prediction method with dynamically changing weight factors, and its formula is as follows.

[0060]

[0061] Indicates that the improved exponential smoothing forecasting algorithm is used to predict t n Compared with the traditional method, the value of the weight factor α in the improved method will increase with the change of t n It changes dynamically with the changes in . It is no longer a fixed value, but is dynamically modified over time.

[0062] (3) For content with a long history of data, i.e. old content, a dynamically optimized LSTM is used for prediction, and the time window size, the number of dense layers, the number of LSTM network layers and the number of neurons are used as optimization objects. m time window candidates, k dense layer candidates, n LSTM layer candidates and u neuron candidates are set, and the mean percentage error (MAPE) of the LSTM model validation data obtained after the number of training times reaches the limit is selected to measure the quality of the parameters. The definition of MAPE is as follows:

[0063]

[0064] y′ n is the value predicted by the model, y n is the true value of the sample, and N is the number of samples. Let M(m,k,n,u) represent the average percentage error obtained after prediction by the LSTM model constructed with the current parameters. Through continuous iterative training, we can select a set of parameters with the best performance of M(m,k,n,u) in the test set, and then determine the model used for old content prediction. At the same time, the optimal m time window values ​​obtained are the standard LB for dividing new and old content.

[0065] Furthermore, in step S3, calculating the total revenue of the content value includes the following steps:

[0066] (1) When the terminal requests content C iIf the content has been cached in the edge server, the edge server can directly satisfy the request and start transmitting the content. The total delay cost at this time is the delay of the upload request L u,i Plus the delay L of downloading content from the server d,i , which can be expressed as:

[0067] L edge,i =L u,i +L d,i

[0068] (2) If the edge server does not cache the requested content, it can only forward the request to the cloud data center. The total latency overhead can be expressed as:

[0069] L cloud,i =L u,i +L etc +L c,i

[0070] L etc is the latency cost of the edge server sending the request to the cloud, L c,i It is the latency overhead required to return content from the cloud to the terminal.

[0071] (3) If a certain content C i Originally, it needs to be obtained from the cloud, with a latency overhead of L cloud,i At this time, putting it in the edge server will shorten the delay cost, and the delay cost will become L edge,i Therefore, the delay gain is defined as follows:

[0072] ΔL i =L cloud,i -L edge,i

[0073] (4) Combined with the content popularity P obtained previously i , together with the delay gain, determine the total value of the content. The specific formula is:

[0074] G i =A*ΔL i +B*P i

[0075] Among them, A and B are two variable factors A+B=1, which is used to control the proportion of the two types of income in the total value income; ΔL i Represents the latency gain of obtaining content, P i represents the popularity gain of the content, G i It represents the total value of the content.

[0076] Furthermore, in step S4, according to the total content value benefit, the cache placement problem is transformed into a benefit value maximization problem, and the cache placement algorithm steps adopted are as follows:

[0077] (1) The original problem makes a decision for each content, 0 represents no choice, 1 represents choice, which is a 0-1 variable, so it is a constrained 0-1 programming problem, which is an NP-hard problem. By considering the content that cache can bring the greatest benefit to the server cluster, the problem is transformed into a benefit maximization problem, and the problem is transformed into the following formula.

[0078]

[0079] G = {G1, G2, ..., Gn} is the calculated total value of the content. The above formula must meet the following restrictions, where S e is the edge server capacity, ca i is a 0-1 variable, ca i =1 means to select content cache, otherwise not cache. i is the size of the content:

[0080]

[0081] (2) The cache placement algorithm adopted by the present invention solves the cache placement problem of content in the server in the following steps:

[0082] 1) Define a two-dimensional array F[i][j] to store the state. F[i][j] represents the maximum benefit that can be obtained by selecting a number of contents from the first i contents and putting them into a server cluster with j remaining spaces.

[0083] 2) Start from 0 and gradually increase the selected capacity. If the current size j is smaller than the content size S i , then the two-digit array F[i][j]=F[i-1][j].

[0084] 3) If S i ≥j, then F[i][j]=max{F[i-1][jS[i]]+Gi,F[i-1][j]}, and the decision of whether to release or not is made based on the revenue of the i-th content.

[0085] 4) Finally, the cache decision set is obtained through reverse deduction.

[0086] The edge server selects the content with the largest total gain from the calculated set for caching. This edge caching method based on content popularity can not only reduce the prediction result error, but also cache the content with larger total gain in the edge server cluster, improve the hit rate of users to obtain content, and reduce the experimental cost of obtaining content.

[0087] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0088] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0089] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0090] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. An edge caching method based on content popularity, characterized in that: It includes the following steps: S1. Build an edge-cloud-end system model based on the server cluster, devices, and user request content, obtain the number of requests for each piece of content at different time periods, and regard the number of requests for the content by end-users per unit time as the historical popularity of the content in that unit time period; S2. Classify historical data of different lengths, and after classification, use an improved exponential smoothing prediction algorithm and a dynamically optimized long short-term memory algorithm to predict the popularity at future moments; S3. Regard the popularity predicted in step S2 as the popularity gain. Since the latency of obtaining content is related to the size of the content itself, if the content is cached in the edge server cluster, the reduced latency compared to obtaining it from the cloud is regarded as the latency overhead gain, and calculate the total value gain of each piece of content; S4. For each server cluster, use a cache placement algorithm to calculate the content sequence with the highest benefit under the current capacity, and cache the content corresponding to this batch of content sequences in the server; S5. Obtain the result of whether each piece of content is cached in the edge server cluster; In step S2, when classifying historical data of different lengths and using an improved exponential smoothing prediction algorithm and a dynamically optimized long short-term memory algorithm to predict the popularity at future moments, it specifically includes: Set lp as the length of the content set and set the length limit LB; (1) If lp < LB, it is classified as new content and predicted using an improved exponential smoothing prediction algorithm; (2) If lp ≥ LB, use a dynamically optimized long short-term memory algorithm to predict the popularity; In step (1), an improved exponential smoothing prediction algorithm is used for prediction; in the traditional exponential smoothing prediction method, the weight factor α determines the size of the final predicted value. When α is larger, the popularity closer to the current time period has a greater impact on the prediction result. Conversely, when the value of α is smaller, the popularity of older content has a greater impact. The formula is as follows; Indicates that the exponential smoothing forecasting algorithm is used to predict t n The popularity of the content at the moment, where Represents the popularity of the content at the last moment. Subsequent operations need to be performed based on the popularity at the last moment. n Represents the current time period, t s represents the starting time period, α is a variable factor, fixed and unchanged, with a value range of [0,1], which is used to determine the influence of the popularity of the previous time period on the future time period; since the weight factor α in the traditional method is static, it is difficult for the smooth prediction to conform to the changes of the time series itself, so the exponential smoothing prediction method with dynamic weight factor change is used, and its formula is as follows; Indicates that the improved exponential smoothing forecasting algorithm is used to predict t n Compared with the traditional method, the value of the weight factor α in the improved method will increase with the change of t n It changes dynamically with the change of , is no longer a fixed value, and is dynamically modified over time, making the accuracy of the prediction closer to the true value; In step (2), using a dynamically optimized long short-term memory algorithm to predict the popularity specifically includes: For content with longer historical data, that is, old content, use a dynamically optimized LSTM for prediction, and regard the time window size, the number of dense layers, the number of LSTM network layers, and the number of neurons as the objects to be optimized; set m candidate values for the time window size, k candidate values for the number of dense layers, n candidate values for the number of lstm layers, and u candidate values for the number of neurons, and select the average percentage error MAPE of the validation data of the LSTM model obtained after the number of training times reaches the limit to measure the quality of the parameters; where the definition of MAPE is as follows: y′ n is the value predicted by the model, y n is the true value of the sample, N is the number of samples, and M(m,k,n,u) represents the average percentage error obtained after prediction by the LSTM model constructed with the current parameters. Through continuous iterative training, a set of parameters with the best performance of M(m,k,n,u) in the test set is selected, and then the model used for old content prediction is determined; at the same time, the optimal m time window values ​​obtained are the standard LB for dividing new and old content.

2. The edge caching method based on content popularity according to claim 1, characterized in that: In step S1, the method of building an edge-cloud-end system model and obtaining historical popularity specifically includes: Most content is stored in the cloud, and the edge server cluster selects some content for storage. The terminal user initiates the request, so the three environments are divided into cloud, edge and terminal. The cloud server is a device far away from the user terminal. Each edge server cluster serves some terminal devices. The terminal devices will initiate requests for certain content in different time periods. The edge server collects historical data of the requested content and obtains the content popularity in different time periods, thereby obtaining the historical popularity set of all content under the edge server cluster for subsequent popularity prediction.

3. The edge caching method based on content popularity according to claim 1, characterized in that: In step S3, the popularity predicted in step S2 is regarded as popularity gain. Since the delay of obtaining content is related to the size of the content itself, if the content is cached in the edge server cluster, the reduced delay compared to obtaining from the cloud is regarded as delay cost gain, and the total value benefit of each content is calculated; The specific formula is shown in the formula; G i =A*ΔL i +B*P i Among them, A and B are two variable factors A+B=1, which is used to control the proportion of the two types of income in the total value income; ΔL i Represents the latency gain of obtaining content, P i represents the popularity gain of the content, G i It represents the total value of the content.

4. The edge caching method based on content popularity according to claim 3, characterized in that: In step S4, the cache placement problem is transformed into a benefit value maximization problem. The original problem makes a decision for each content, 0 represents no selection, 1 represents selection, which is a 0-1 variable, and it is a constrained 0-1 planning problem, that is, an NP-hard problem. By considering the content that the cache can bring the greatest benefit to the server cluster, the problem is transformed into a benefit maximization problem, and the problem is transformed into the following formula: That is, the content with the largest total gain is selected from the collection for caching.

Images