A mobile edge cache optimization method based on federated learning

Abstract

The present invention relates to a mobile edge cache optimization method based on effective federated learning, and belongs to the fields of Internet of Things and artificial intelligence. The method takes into account the changing user mobility and content popularity within a single base station, and places the requested content in the edge cache in advance by predicting the content popularity to improve the cache hit rate. Specifically, the RWP random path point model is used to obtain the user time trajectory table to simulate the user's mobile path. Taking into account the local training consumption, the users who participate in the FL local training are selected by combining clustering with thresholds, and the model weights are controlled by the attention mechanism to perform global model aggregation. According to the obtained global prediction model, the predicted request content is cached to the server in advance to improve the cache hit rate. The method uses a federated learning method to optimize client selection and weight aggregation, and implements an effective federated learning method to reduce local training consumption and improve cache hit rate.

Description

Technical Field

[0001] This invention belongs to the fields of Internet of Things and artificial intelligence, and specifically relates to a mobile edge caching optimization method based on federated learning. Background Art

[0002] With the development of the Internet of Things (IoT) and communication technologies, the number of next-generation IoT devices is exploding, and users have increasingly higher latency requirements for content. Retrieving content from remote data centers may not meet these low-latency needs. Mobile Edge Computing (MEC) brings computing and caching services from mobile network operators (MNOs) or the cloud to the network edge (e.g., base stations). Content can be cached at the edge to improve Quality of Experience (QoE). Therefore, edge caching research has become one of the hottest research topics in the field of wireless communication. Since edge nodes have limited storage space, caching schemes need to identify and cache popular content that is of interest to most users. Based on content popularity, caching popular content on the edge cloud can reduce network link congestion and request latency, thereby improving the user's QoE.

[0003] Traditional intelligent algorithms for predicting content popularity require sending user data to a central server for training. This transmission of large amounts of unprocessed data puts pressure on the network and threatens user privacy. In 2016, Google proposed Federated Learning (FL), a mobile edge computing framework for machine learning (ML) technology. It utilizes distributed client data and computing resources to train high-performance ML models while protecting client data privacy. In Federated Learning, data on terminal devices is not independent and identically distributed, making it suitable for the independence of terminal devices in MEC environments and easily scalable.

[0004] By using federated learning to design edge caching methods, we can predict content popularity and cache popular content while protecting client data privacy, thereby reducing network link congestion and request latency. For example, Yu et al. proposed a federated learning-based proactive content caching (FPCC) method. Based on a hierarchical architecture, each user trains their local data using a hybrid filtering method on a stacked autoencoder, predicts content popularity based on the local data, and then the server averages the model parameters from the user's end to build a global model, selecting the most popular files for edge caching. Jiang et al. proposed a novel context-aware popularity prediction strategy based on federated learning. By considering user preference learning and using contextual information for clustering, they formulated an effective global popularity prediction model based on local model learning. Gao et al. proposed a caching strategy that considers mobility and content caching popularity using federated learning. By considering the concept of One-Stop Service (OSS), they proposed predicting the density of pedestrians within the range of each base station and using the prediction results to calculate the request probability of a single file. By establishing a minimization model that minimizes routing cost and file request probability, they used a greedy algorithm to select the set of content to be cached while minimizing routing cost. Yu et al. proposed a federated learning-based mobile-aware proactive edge caching scheme (MPCF), which uses a context-aware adversarial autoencoder to predict highly dynamic content popularity. Although the above schemes use different methods to predict content popularity, there are still some problems in the MEC environment, even within the range of small base stations, due to the high density and mobility of terminal devices: (1) All users within the range of the base station participate in training, causing unnecessary consumption of terminal devices; (2) Due to the different computing power and data quality of terminal devices, the quality of the local model obtained is uneven. Simply averaging and aggregating to obtain the global model will slow down the convergence speed of the global model.

[0005] To address the aforementioned issues, the present invention aims to propose an edge cache optimization method based on effective federated learning, taking into account user mobility. By optimizing user selection and model aggregation in federated learning, this method aims to accelerate global model convergence and improve cache hit rate while reducing local user consumption. Summary of the Invention

[0006] This invention addresses the issue of improving cache hit rates by predicting content popularity and placing cached data accordingly. In a MEC (Multi-access Edge Computing) environment, centralized training of all user data raises concerns about user data privacy and network load due to data transmission issues. To resolve this, this invention employs a federated learning approach. This involves placing the same prediction model on both the user and MEC servers. User data trains the local model, and the server aggregates the parameters of the local model to obtain a global model for predicting content popularity. To achieve this, the invention proposes a user selection and aggregation method during the federated learning process. This reduces unnecessary terminal device consumption caused by all users participating in training within the same base station range, while simultaneously improving cache hit rates and reducing user request latency.

[0007] To address the aforementioned issues, this invention proposes a mobile edge caching optimization method based on federated learning. Considering user mobility, the method utilizes federated learning to train a content popularity prediction model. By selecting users participating in federated learning training and optimizing the global model aggregation method, it aims to predict content popularity and proactively cache at the edge while reducing terminal device consumption, thereby improving cache hit rate and reducing user request latency. To achieve this goal, firstly, a user mobility model needs to be established to determine the user's mobility range. Secondly, during federated training, the similarity of requested content among users is utilized using the DBSCAN (Density Based Spatial Clustering of Applications with Noise) clustering algorithm. A threshold is introduced to select a certain proportion of users from the cluster to participate in training. Then, an attention mechanism is used to control the model aggregation weights to balance the uneven contribution of users' local model quality to the global model. The model is trained in a distributed manner, and a global model of users within the base station range is obtained through weight aggregation. Based on the obtained global prediction model, the predicted request content is proactively cached at the edge in advance to accelerate global model convergence and improve cache hit rate while reducing local consumption.

[0008] The specific technical solution is as follows:

[0009] A mobile edge caching optimization method based on federated learning. This method considers the constantly changing user mobility and content popularity within a single base station range, and improves cache hit rate by predicting content popularity and placing it in advance in the edge cache. The federated learning (FL) method is used to predict content popularity. The same prediction model is deployed on each user and the MEC server. The local prediction model is based on the user's historical request data, and the global model is obtained by aggregating user model parameters on the server. The key feature is that the user's time-trajectory table is obtained using the random path point (RWP) model to simulate the user's movement path. Considering the local training cost, users participating in the local FL training are selected through a combination of clustering and thresholding. An attention mechanism is used to control the model weights for global model aggregation. Based on the obtained global prediction model, the predicted request content is cached on the server in advance to improve the cache hit rate. The main steps include:

[0010] Step 1. Use the RWP (Random Path Point) model to obtain the user time trajectory table, and obtain the user set within the base station coverage area of ​​each time slot. The user set U within the base station coverage area at time slot t is... t The method to obtain it is as follows:

[0011] The time period is divided into T segments using the concept of time slots, denoted as Γ = {1, 2, ..., t, ..., T}. The user u generated using the RWP model... i At time slice t, calculate user u i The distance from the base station is used to determine if the user is within the base station's coverage area at time t. If this distance is less than the base station's coverage area, then the user is within the base station's coverage area at time t. This process is repeated for all users to obtain the user set U within the base station's coverage area at time t. t .

[0012] Step 2. Based on the user set U obtained in Step 1 t Further selection of users to participate in local FL training is performed using the following method.

[0013] Step 2.1) Calculate the set U of users who have moved to the base station's range at time t. t The preferences of all users, where the i-th user u i preferences User u equals t-1 i Model parameters Compared with the user's historical request data at time t-1 Dot product; send preferences to the MEC server to calculate the cosine similarity of preferences among any users in the user set, where user u i With user u j The specific formula for the cosine similarity of preferences is as follows:

[0014]

[0015] The DBSCAN clustering algorithm is used to cluster users by taking the cosine similarity of preferences among users as input.

[0016] Step 2.2). Let the clustering result be class a, and the labels be {1, 2, ..., a-1, None}, where the label None represents users whose similarity to other users is low due to the uncertainty of user movement location and requested content, and the number of users in each cluster is equal to 1. The number of users in each class is {N1, N2, ..., N}. a N none}, N i This represents the number of users in one category, where N is the sum of the number of users not labeled "None". sel ,Right now According to the threshold T max and T min The proportion of each user group participating in training is selected as follows:

[0017]

[0018] in This represents the proportion of users in the None category out of all user categories. pro This represents the selection ratio of the None class, that is, randomly selecting N from the None class. none *N pro N users participated in the training. sel To determine the selection ratio for non-None classes, select N from each class. i *N sel There are 3 users, where i = {1, 2, ..., a-1}, and p and q are random values ​​in the range (0.2, 0.5); the set of users selected in each class is the user set C for segment t participating in model training. t C t ∈U t .

[0019] Step 3. Train the FL local model on the users selected in Step 2 in a distributed manner. Aggregate the data by controlling the aggregation weights to obtain the global model. Based on the obtained global model, cache the predicted popular content on the server. The specific process is as follows:

[0020] Step 3.1). The MEC server will use the FL global model parameters w obtained at time t-1. t-1 The user set C sent to fragment t for model training t All users in the dataset are used as the initial parameters of the user model at time t; user u i Use the local historical data corresponding to fragment t The trained model receives updated user model parameters. The average loss value is The number of batches required to complete one epoch is For fragment t, user u i The proportion of data used to train the model to the total data of all users participating in the training.

[0021] Step 3.2). The percentage of users who participate in the training will receive Model parameters Average loss as well as The data is uploaded to the server for aggregation, and the user's aggregation weight is obtained based on an attention mechanism. Aggregate weight vectors for users, where parameter and User set C t All users The vector formed This represents the model's objective, namely, minimizing the average loss, maximizing the proportion of data used in training, and maximizing the number of batches. For vectors Dimensions With Q t They have the same dimensions;

[0022] The global model f(w) on the server side at time t t The following is represented:

[0023]

[0024]

[0025] Among them, w t The parameters of the global model on the server side when t is . For user u i Aggregated weights of model parameters For local user models, For the i-th user u i The local target loss function, x i，k This represents the k-th local training data of the i-th user.

[0026] Based on the global prediction model obtained in fragment t, the popularity of the content is predicted, and the resulting set of popular content files is placed in the cache to meet the user's request at fragment t, thereby improving the cache hit rate.

[0027] Beneficial effects

[0028] The mobile edge caching framework based on federated learning designed in this invention uses user similarity for clustering, selects users to participate in local FL training according to the proposed user selection method, and uses an attention mechanism to control the aggregation weights. This accelerates model convergence and improves cache hit rate while reducing training costs. Attached Figure Description

[0029] Figure 1 Example diagram of system model

[0030] Figure 2 A diagram illustrating the categories based on DBSCAN clustering.

[0031] Figure 3 The proposed federated learning communication framework diagram is trained throughout the t-round training process.

[0032] Figure 4 A comparison chart of the training loss of the proposed method and FPCC.

[0033] Figure 5 The graph shows a comparison of cache hit rates between the proposed method and other caching algorithms at different cache sizes. DETAILED DESCRIPTION

[0034] To make the objectives, technical solutions, and implementation of the present invention clearer, the following detailed description of the invention will be provided in conjunction with the accompanying drawings.

[0035] This invention relates to a mobile edge caching optimization method based on federated learning. In the mobile edge caching architecture, the same prediction model architecture is deployed on the MEC server and users, and federated learning is used to train the prediction model. Users train their models based on their local request datasets, and the global model on the server side is obtained by aggregating the model parameters of local users. Users provide the server with relevant parameters of their user models, which serve as the basis for selecting users to participate in training and aggregating user model parameters. The popular file set of predicted content is placed in the cache based on the obtained global model to improve the user's hit rate and meet user needs.

[0036] The specific steps are as follows:

[0037] Step 1. Based on the cache optimization under the consideration of user mobility, use the RWP random path point model to obtain the user time trajectory table, and obtain the user set within the base station coverage area of ​​each time slice.

[0038] The entire system is divided into several regions, with multiple users distributed within the system. Each region is equipped with a base station and a nearby MEC server. Users are provided with services by the base station through a wireless network link, and users have a certain computing power. Figure 1This is an example diagram of the system model. The system range on the coordinate axis is set to (-2v, 2v, 4v). Specifically, it represents a square region with a side length of 4v and a top-left point of (-2v, 2v). Within this region, there are a total of N users, i.e., U = {u1, u2, ..., u...}. N The user moves throughout the entire system; the concept of time slots is used to divide the considered time period into T segments, represented as Γ={1, 2, ..., t, ..., T}. This represents the local request history dataset for each user in segment t. This represents the number of data sets, specifically the number of the i-th user u. i Dataset D i The quantity is The base station BS is placed at a coordinate axis distance of v meters from the origin. The location of the base station is (0, v). The effective range of the base station is smaller than the system range.

[0039] Step 1.2). Use the RWP model to generate the motion trajectories of all users within the system, obtaining a time-period trajectory table for all users, where user u i The position at time t is denoted as Calculate u for each user at time t i The distance to the base station is calculated using the following formula:

[0040]

[0041] like Indicates user u i If the user is within the base station's coverage area at time t, then the opposite applies, yielding the user set U within the base station's coverage area at time t. t .

[0042] Step 2. A mobile edge caching optimization method based on federated learning, utilizing the entire process of user selection and model aggregation using federated learning;

[0043] S1. In the initial stage t=1, the user set U within the current time range of the base station is... t The user group consists of all users participating in model training, and the i-th user u i With local historical request dataset (This is the initial dataset) used as input to the model to obtain the model's update parameters. Users upload model parameters to the MEC server. The server then aggregates the average of the user model parameters used in training to obtain the FL global model, which is the aggregated weight p of the model. i The mean of all user model parameters is used to obtain the possible set of files requested by users based on the global prediction model. The MEC server then places the set of files into the cache.

[0044] S2. When t > 1, calculate the set U of users who have moved to the base station's range at time t. t The preferences of all users, where the i-th user u i preferences User u equals t-1 i Model parameters Compared with the user's historical request data at time t-1 Dot product; send preferences to the MEC server, where the server calculates the cosine similarity of preferences among any users in the user set, where user u i With user u j The specific formula for the cosine similarity of preferences is as follows: The DBSCAN clustering algorithm is used to cluster users by taking the cosine similarity of preferences among users as input.

[0045] S4. Select users to participate in training based on the clustering results;

[0046] Let the clustering result be class a, and the labels be {1, 2, ..., a-1, None}, where the label None represents users whose similarity to other users is low due to the uncertainty of user movement location and requested content, and the number of users in each cluster is equal to 1. The number of users in each class is {N1, N2, ..., N}. a N none}, N i This represents the number of users in one category, where N is the sum of the number of users not labeled "None". sel ,Right now According to the threshold T max and T min The proportion of each user group participating in training is selected as follows:

[0047]

[0048] in This represents the proportion of users in the None category out of all user categories. pro This represents the selection ratio of the None class, that is, randomly selecting N from the None class. none *N pro N users participated in the training. sel To determine the selection ratio for non-None classes, select N from each class. i *N sel There are 3 users, where i = {1, 2, ..., a-1}, and p and q are random values ​​in the range (0.2, 0.5); the set of users selected from each class is the user set C for segment t participating in model training. t C t ∈U t ;

[0049] The S5.MEC server will use the FL global model parameters w obtained at time t-1. t-1 The user set C sent to fragment t for model training t All users in the dataset are used as the initial parameters of the user model at time t; user u i Use the local historical data corresponding to fragment t The trained model receives updated user model parameters. The average loss value is The number of batches required to complete one epoch is For fragment t, user u i The proportion of data used to train the model to the total data of all users participating in the training.

[0050] S6. The percentage of users who will receive training Model parameters Average loss as well as The data is uploaded to the server for aggregation, and the user's aggregation weight is obtained based on an attention mechanism. Aggregate weight vectors for users, where parameter and User set C t All users The vector formed This represents the model's objective, namely, minimizing the average loss, maximizing the proportion of data used in training, and maximizing the number of batches. For vectors Dimensions With Q t They have the same dimensions;

[0051] Instruction manual attached Figure 2The proposed federated learning communication framework (FLCM) is illustrated in step 3 (S2-S6) at time t. First, all users within the base station's range (u1, u2, u3 in the diagram) send their local preferences to the MEC platform. At the MEC, user similarity is calculated and clustered. A threshold selection strategy is proposed to select users to participate in training, and the current global model parameters are sent to the selected users as the initial parameters for the user models at time t. (u1 and u3 in the diagram represent the selected users at time t). The local FLCM models are fitted and trained based on the user's historical request dataset, and the recorded user model parameters and user-related information are sent to the server. The server uses an attention mechanism to obtain the weights for each user model. The MEC aggregates all local parameters based on these weights to obtain the global model, completing one global model training iteration. The resulting dataset is then stored on the current MEC server.

[0052] S6. The server obtains the most popular files requested by users within the current coverage area based on the global prediction model and puts them into the cache to meet the needs of users in the current segment.

[0053] S7. Repeat steps two through six until segment T is reached.

[0054] Through simulation Figure 3 The proportion of locally generated training data and FPCC in the algorithms proposed in different segments shows that, across the simulations, the highest amount of locally generated training data reached 75% of the FPCC. From... Figure 4 It can be seen that compared to the FPCC algorithm, the proposed algorithm achieves better model accuracy; and from Figure 5 It can be seen that as the cache size increases, the proposed algorithm exhibits a better hit rate than FPCC and random. This invention demonstrates that, with a maximum training data size of only 75% of FPCC, it achieves a higher hit rate than FPCC, and also achieves rapid convergence.

Claims

1. A mobile edge cache optimization method based on federated learning, characterized in that: This method takes into account the changing user mobility and content popularity within a single base station, predicts content popularity through the federated learning FL method, and places popular content in the edge cache in advance to improve the cache hit rate; the same prediction model is deployed on each user and MEC server, the user-side model is called the local model, and the server-side model is called the global model. The local model is trained based on the user's historical request data, and the global model is obtained by the server aggregating the user-side model parameters. The user selection method and the global model aggregation method include the following steps: Step 1), use the RWP random path point model to obtain the user time trajectory table, and obtain the user set within the coverage range of each time slice base station. There are N users in total, among which U t It represents the set of users within the coverage area of ​​the base station at time slice t; Step 2), select users to participate in FL local training; Step 3) Train the FL local model for the user selected in step 2) in a distributed manner, upload the obtained model parameters to the server, use the attention mechanism to control the model weights for aggregation, and obtain the global model. According to the obtained global model, cache the predicted popular content to the server, and then make user requests to improve the cache hit rate; Furthermore, the aggregation process of the FL global model at time slice t specifically includes the following steps: Step 3.1). The MEC server takes the FL global model parameter w obtained at time t-1 t-1 The user set C that participates in model training at time slice t is sent t All users in the , as the initial parameters of the user model at time slice t; user u i Use the local historical data corresponding to time slice t Train the model to get updated user model parameters The average loss is The number of batches required to complete an epoch is For time slice t, user u i The proportion of data for training the model to all user data involved in training in, is user u at time slice t i The amount of data involved in training; Step 3.2) The percentage of users who participated in the training will receive Model parameters Average loss as well as Uploaded to the server for aggregation, the user's aggregation weight is obtained according to the attention mechanism, that is, is the user aggregation weight vector, where parameter and The user set C t All users in The vector composed of Represents the goal of the model, namely the minimum average loss, the maximum proportion of training data, and the maximum number of batches. For vector The dimension of With Q t have the same dimensions; The global model f(w t ) is expressed as follows: Among them, w t are the parameters of the global model on the server at time slice t, For user u i The aggregate weights of the model parameters, For the local user model, For the i-th user u i The local objective loss function, x i,k represents the kth local training data of the ith user.

2. The mobile edge cache optimization method based on federated learning according to claim 1, characterized in that: The user set U within the coverage area of ​​the base station at time slice t t The method to obtain is as follows: The concept of time slots is used to divide the considered time period into T segments, represented as Γ = {1, 2, .., t, …, T}, and the RWP model is used to generate user u i At the time slice t, calculate the position of user u i The distance from the base station. If the distance is less than the base station coverage, then the user is within the base station coverage at time slice t. Traverse all users to obtain the user set U within the base station coverage at time slice t. t .

3. The mobile edge cache optimization method based on federated learning according to claim 2, characterized in that: Furthermore, the method for selecting users to participate in FL local training at time slice t is as follows: Step 2.1) Calculate the user set U that moves into the range of the base station at the current time slice t t The preferences of all users in , where the i-th user u i Preference When user u is equal to t-1 i Model parameters The historical request data of the user at time t-1 Point product; send the preference to the MEC server and calculate the cosine similarity of the preference between any users in the user set, where user u i With user u j The specific formula for the preferred cosine similarity is: Using the DBSCAN clustering algorithm, the cosine similarity of preferences between users is used as input to cluster users; Step 2.2). Assume that the clustering result is a class, and the labels are {1,2,…,a-1,None}, where the label None indicates that due to the uncertainty of the user's mobile location and request content, the similarity with other users is low, and the number of clusters is equal to 1. The number of users in each class is {N1,N2,…,N a ,N none },N i It is represented by the number of users in one category. The sum of all non-None label users is N. sel ,Right now According to the threshold T max and T min Select the proportion of each type of user participating in training, as follows: in Indicates the ratio of None users to all users, N pro It is expressed as the selection ratio of the None class, that is, N is randomly selected from the None class. none *N pro Users participate in training, N sel is the selection ratio of non-None classes, N is selected in each class i *N sel users, where i = {1, 2, ..., a-1}, p and q are random values ​​in the range of (0.2, 0.5); the set of users selected in each category is the user set C participating in model training at time slice t t , where C t ∈U t .

Images