Active edge caching method and system based on multi-scale feature fusion
By integrating mobile edge computing and multi-scale feature fusion in the social content center network, using the multi-head attention mechanism to predict the future popularity of content, the problems of low cache hit rate and high complexity in the existing technology are solved, and efficient cache is achieved Strategy improves user experience.
Patent Information
- Application Number
- CN202411022015.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-22
AI Technical Summary
The existing active caching strategy has low cache hit rate in social networks, unable to fully utilize network resources, and has high complexity, unable to implement parallel computing, unable to effectively predict the future popularity of content, and affecting user experience.
By building a system network model, the mobile edge computing server is integrated into the social content center network, and multi-scale features of historical popularity, social relationships and user mobility are obtained, and the history of content is extracted using multi-head attention mechanism and two-way long and short-term memory networks. and future timing characteristics, predict the future popularity of content, and cache popular content to edge servers near users in advance.
It improves the cache hit rate, reduces the overall system content access latency, improves the overall caching efficiency of the system, improves the user experience quality, and is better than traditional caching methods.
Smart Images

Figure CN119011670A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of active caching, and in particular relates to an active edge caching method and system based on multi-scale feature fusion. Background Art
[0002] In recent years, social networks have become an indispensable part of people's daily lives, and their traffic accounts for a growing proportion of mobile data traffic. To meet this challenge, information-centric networks have been proposed as a new type of network architecture, of which content-centric networks are one of its most promising branches. Its core feature is the ubiquitous cache within the network, which can effectively reduce the latency of users accessing content and improve the quality of user experience. Applying the content-centric network architecture to social networks has formed a social content-centric network, which can effectively improve the efficiency of content transmission in social networks and reduce network congestion.
[0003] Active caching in edge networks occurs before users make requests. It predicts the popularity of content in the future by analyzing the content's identity, type, and popularity, and caches popular content in base stations near users in advance. Therefore, designing an accurate content popularity prediction method to achieve an efficient caching strategy is the focus of existing research.
[0004] Since the popularity of content in social networks is dynamic and is affected by many factors, such as time, user interests, and events, it is very difficult to predict the future popularity of content. For existing active caching strategies, most of them only consider the historical content popularity, while ignoring multi-dimensional features such as future content popularity, user social relationships, and user mobility, resulting in a low cache hit rate and inability to fully utilize network resources. In addition, most social content-centric network caching strategies usually use long short-term memory networks, which are highly complex and cannot achieve parallel computing during the training phase. The attention mechanism has become a popular optimization method for sequence modeling, which can model without considering the distance of model data dependencies in the sequence. Summary of the invention
[0005] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides an active edge caching method and system based on multi-scale feature fusion. By considering the historical popularity, social relationships, geographic location and future popularity of different content for multi-scale feature fusion, the future popularity of the content can be effectively predicted and the popular content can be cached in advance to the edge server near the user, thereby improving the cache hit rate and reducing the overall content access delay of the system, thereby improving the overall cache efficiency of the system.
[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:
[0007] A first aspect of the present invention provides an active edge caching method based on multi-scale feature fusion.
[0008] The active edge caching method based on multi-scale feature fusion includes the following steps:
[0009] Build a system network model to integrate the mobile edge computing server into the social content center network;
[0010] Based on the mobile edge computing server, multi-scale features representing the historical popularity of different contents, user social relationships and user mobility are obtained respectively, and the multi-scale features are fused to obtain a unified time series feature representation;
[0011] The unified time series feature representation is input into the encoder to extract the historical and future time series features of the content. The output vector of the middle layer of the encoder is input into the decoder. The predicted popularity values for multiple time periods in the future are obtained based on the multi-head attention mechanism. The average value of the predicted popularity values for multiple time periods is taken as the final prediction result.
[0012] The content with the highest ranking in the final prediction results is used as the content to be cached and cached in advance.
[0013] A second aspect of the present invention provides an active edge caching system based on multi-scale feature fusion.
[0014] Active edge caching system based on multi-scale feature fusion, including:
[0015] The integration module is configured to: construct a system network model to integrate the mobile edge computing server into the social content center network;
[0016] The multi-scale feature fusion module is configured to: based on the mobile edge computing server, respectively obtain multi-scale features representing the historical popularity of different contents, user social relationships and user mobility, and fuse the multi-scale features to obtain a unified time series feature representation;
[0017] The prediction module is configured to: input the unified time series feature representation into the encoder, extract the historical and future time series features of the content, and input the intermediate layer output vector of the encoder into the decoder, obtain the predicted popularity values of multiple time periods in the future based on the multi-head attention mechanism, and take the average value of the predicted popularity values of multiple time periods as the final prediction result;
[0018] The cache module is configured to: cache the content with the highest final prediction result as the content to be cached in advance.
[0019] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the active edge caching method based on multi-scale feature fusion as described in the first aspect of the present invention.
[0020] A fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the active edge caching method based on multi-scale feature fusion as described in the first aspect of the present invention are implemented.
[0021] One or more of the above technical solutions have the following beneficial effects:
[0022] The present invention provides an active edge caching method and system based on multi-scale feature fusion, which aims to improve the cache hit rate and reduce the cache content redundancy, thereby improving the performance and user experience of the social content-centric network. Specifically, the present invention first integrates mobile edge computing into the social content-centric network to enhance computing power and increase the network edge cache space; secondly, a content popularity prediction model based on a multi-head attention mechanism is designed, which integrates multi-dimensional features including historical popularity information, user social relationships, geographic location and future popularity information to accurately predict content popularity, and adopts an encoder-decoder structure with a bidirectional long short-term memory network and a multi-head attention mechanism, and uses multi-dimensional historical features to perform multi-step popularity prediction; finally, the average value of the multi-step predicted popularity is used as the final prediction result, and a cache placement and replacement strategy is designed based on content popularity and node distance. Overall, the present invention can improve the cache hit rate and reduce the overall content access delay of the system, thereby improving the overall cache efficiency of the system.
[0023] The cache strategy designed by the present invention was compared with four baseline strategies (History Popularity-basedPredict, DNN-based Predict, Random and Optimal). The four baseline schemes are a cache strategy that only considers the historical popularity of the content and ignores multi-dimensional features, a cache strategy based on a deep neural network, a random content cache strategy, and an optimal cache strategy. Experiments have shown that the cache strategy proposed by the present invention is superior to the above four baseline strategies in terms of content popularity prediction accuracy, system average content access delay, and cache hit rate, and has better cache utilization and higher user experience quality.
[0024] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] The accompanying drawings in the specification, which constitute a part of the present invention, 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 improper limitations on the present invention.
[0026] Figure 1 It is a schematic diagram of the system model architecture of the present invention.
[0027] Figure 2 It is a schematic diagram of the multi-scale feature fusion implementation process of the present invention.
[0028] Figure 3 It is a flow chart of the content popularity prediction implementation and multi-head attention mechanism implementation of the present invention. DETAILED DESCRIPTION
[0029] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0030] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.
[0031] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.
[0032] Embodiment 1
[0033] This embodiment discloses a method for placing edge cache based on a multi-scale feature fusion and a multi-head attention mechanism. Aiming at the necessity of predicting content popularity in social content-centric networks and the problem of low cache hit rate caused by the lack of multi-dimensional feature consideration in existing cache strategies, the method integrates mobile edge computing into social content-centric networks, designs a method for predicting content popularity based on a multi-scale feature fusion and a multi-head attention mechanism, and designs a real-time active edge cache placement and content replacement method. By considering the historical popularity, social relationship, geographic location and future popularity of different contents for multi-scale feature fusion, the future popularity of the content is effectively predicted and the popular content is cached in advance to the edge server near the user, thereby improving the cache hit rate and reducing the overall content access delay of the system, and improving the overall cache efficiency of the system. The beneficial effect of the present invention is that, compared with the traditional cache method, the scheme can significantly improve the content cache hit rate and improve the cache efficiency.
[0034] Specifically, the active edge caching method based on multi-scale feature fusion in this embodiment may include the following steps:
[0035] Build a system network model to integrate the mobile edge computing server into the social content center network;
[0036] Based on the mobile edge computing server, multi-scale features representing the historical popularity of different contents, user social relationships and user mobility are obtained respectively, and the multi-scale features are fused to obtain a unified time series feature representation;
[0037] The unified time series feature representation is input into the encoder to extract the historical and future time series features of the content. The output vector of the middle layer of the encoder is input into the decoder. The predicted popularity values for multiple time periods in the future are obtained based on the multi-head attention mechanism. The average value of the predicted popularity values for multiple time periods is taken as the final prediction result.
[0038] The content with the highest ranking in the final prediction results is used as the content to be cached and cached in advance.
[0039] It can be understood that the following "multi-step" in this embodiment refers to the prediction of content popularity for multiple time periods in the future in the content popularity prediction task. The multi-step popularity prediction values are then averaged to reflect the average popularity of the content at multiple future time points.
[0040] Furthermore, the specific technical solution adopted in this embodiment mainly includes the following steps:
[0041] Step 1: Build a system network model and integrate the mobile edge computing server into the social content center network to enhance computing power and increase network edge cache space.
[0042] Step 2: Construct a multi-scale feature matrix to collect multi-dimensional feature data such as content historical popularity, user social relationships, and user mobility. Use bidirectional long short-term memory networks and graph neural networks to extract features, output feature vectors, and expand them into two-dimensional feature maps. Based on a multi-dimensional feature pyramid network, semantically fuse feature maps of different dimensions and integrate them into a unified time series feature representation.
[0043] Step 3: Predict content popularity based on the multi-head attention mechanism. Input the fused time series features into the bidirectional long short-term memory network (Bi-LSTM) model. The encoder is used to extract the historical and future time series features of the content, and transmit the output vector of the intermediate layer of the Bi-LSTM to the decoder. The decoder inputs the popularity value of the current time step and outputs the predicted popularity value of multiple steps in the future.
[0044] Step 4: Design cache placement and cache replacement strategies based on predicted content popularity. Perform weighting according to content popularity and node distance, calculate the weighted popularity of the content, compare the weighted popularity with the local cache for cache placement, and perform cache replacement based on the predicted popularity of the content and request time.
[0045] The step 1 specifically includes:
[0046] Consider a social content-centric network integrated with mobile edge computing capabilities, where the set of mobile edge servers is Connected to each other through base station links, user devices with mobility and social relationships are distributed around each edge server. Cloud data center provides content library The mobile edge server provides both computing power and content caching. It can complete the calculation of content popularity indicators in a short time and cache the top k popular content in advance to the corresponding base station. When the content requested by the user is in the cache list of the edge server, a cache hit occurs. Otherwise, the requested content will be obtained from the cloud data center.
[0047] The step 2 specifically includes:
[0048] Step 2-1: Extract historical popularity features and define the content historical popularity feature matrix X h , which contains information such as the number of content visits at multiple time steps, user ratings of content, content type, etc., and uses Bi-LSTM to perform a feature matrix X on the content history popularity. h Perform feature extraction to obtain long-distance dependencies in the sequence and output the historical feature vector H = {h1,h2,…,h n}, where n is the number of historical time steps.
[0049] Step 2-2: Define the user social relationship feature matrix X s , which contains information such as users’ social network activities, interaction frequency between users, and social influence. In order to effectively process graph structure data and extract the association between nodes, a graph neural network (GNN) is used to map the social relationship matrix X s For feature extraction, input X s , the output is the user social relationship feature vector S = {s1, s2, ..., s u}, where u is the number of users.
[0050] Step 2-3: Define the user mobility feature matrix X m , including user location information, mobility patterns and other information, collect user mobility trajectory data, that is, user geographic location information in different time periods, calculate the user's moving distance and moving speed in different time periods based on the user's geographic location information in different time periods, combine each user's moving distance and moving speed into a feature vector, and splice all users' mobility feature vectors into a matrix M = {m1,m2,…,m u}, where u is the number of users.
[0051] Step 2-4: Multi-scale feature matrix fusion. Using the multi-dimensional feature pyramid network (MFPN), upsampling and compression operations are used in both spatial and channel dimensions to semantically fuse feature maps of different dimensions and integrate them into a unified feature representation.
[0052] The step 3 specifically includes:
[0053] Step 3-1: Construct a BiLSTM encoder-based and multi-head attention-based decoder structure. The encoder uses the BiLSTM layer to extract historical time series features, and the BiLSTM intermediate layer output vector is input into the decoder, which contains the historical and future trends of content popularity. The decoder inputs the popularity value of the current time step and outputs the predicted popularity value of multiple steps in the future.
[0054] Step 3-2: Use Bi-LSTM to predict the popularity of request content to extract historical and future time series features, considering only the forward direction in the model time step, using the input gate g i , output gate g o 、Forget Gate f And two activation functions sigmod and tanh are used for data processing, and the intermediate layer output vector h t Passed to the multi-head attention module in the decoder.
[0055] Step 3-3: Convert the input vector [h t-τ ,h t-τ+1 ,…,h t ]It is divided into multiple heads, i.e., multiple subspaces, through multiple fully connected layers. Each subspace independently calculates the query matrix Q, key matrix K, and value matrix V, and then independently obtains its own attention score through dot product calculation. Then, the multi-head attention scores are concatenated and projected into the output space.
[0056] Step 3-4: The output of each attention head is further processed through a feedforward neural network. The feedforward neural network consists of two linear transformations and a ReLU activation function:
[0057] FFN(x)=ReLU(xW1+b1)W2+b2
[0058] Among them, W1 and W2 are weight matrices, and b1 and b2 are bias parameters.
[0059] The step 4 specifically includes:
[0060] Step 4-1: Weight the path length according to the content popularity and node distance, and divide it into two segments. The node closer to the request node has a higher cache probability. Calculate the weighted popularity of the content pop weight :
[0061]
[0062] Among them, pop_current represents the current popularity of the content, len_distance represents the total number of hops for content transmission, and len_current is the number of hops from the request node to the current node.
[0063] Step 4-2: Make cache replacement decisions based on the predicted popularity of the content and the request time. The algorithm execution process is as follows:
[0064] S1: When the content reaches the node, calculate the weighted popularity of the content pop weight .
[0065] S2: Clear the timed-out content in the local cache.
[0066] S3: If the local cache is not full, pop the content weight and request time are stored in the cache.
[0067] S4: If the local cache is full, compare the pop of the content weight The content with the least popularity in the cache.
[0068] S5: If the content pops weight If the popularity of the content is greater than the minimum popularity in the cache, the content with the minimum popularity in the cache is deleted and the new content is added to the cache.
[0069] As a more detailed implementation:
[0070] The system model involved in this embodiment is as follows Figure 1 As shown in Figure 2, consider a social content-centric network that integrates mobile edge computing capabilities, where the set of mobile edge servers is Connected to each other through base station links, user devices with mobility and social relationships are distributed around each edge server. Cloud data center provides content library The mobile edge server provides both computing power and content caching. It can complete the calculation of content popularity indicators in a short time and cache the top k popular content in advance to the corresponding base station. When the content requested by the user is in the cache list of the edge server, a cache hit occurs. Otherwise, the requested content will be obtained from the cloud data center.
[0071] Then, if Figure 2As shown in the figure, we collect multi-dimensional feature data such as historical popularity of different contents, user social relationships and user mobility to extract features and obtain feature vectors. User social relationship feature vector User mobility feature vector Expand to a two-dimensional input feature map to get c1, c2, c3, let c = [c1, c2, c3], the size is [[B, C1, H1, W1], [B, C2, H2, W2], [B, C3, H3, W3]], where B, C, H, W represent the batch size, channel size, spatial height and width respectively. The sizes of C, H, W are inconsistent, and 1*1 convolution is used to change their channels to the same channel size.
[0072] Next, a multi-dimensional feature pyramid network is constructed to semantically fuse feature maps of different dimensions into a unified time series feature representation.
[0073] like Figure 2 As shown, first, feature fusion is performed in the channel dimension, and the unfold operation is performed to change the shape of the feature map to [B, C′, L], where the sliding window size C′=C*K*K, and the number of sliding windows L is expressed by the following formula:
[0074]
[0075] Where K is the convolution kernel size, padding is the padding size, stride is the stride size, and then the output c′=[c′1,c′ 21 ,c′ 22 ,c′3] is expressed as:
[0076] c′=f unfold (W a *C)
[0077] Among them, W a represents a 1*1 convolutional layer, f unfold represents the unfold function. Finally, after the reshape function, the channel dimension feature fusion output is p′=[p′1,p′2,p′3], which is expressed as:
[0078] p1′=f reshape (c′1+c′ 21 )
[0079] p′2=f reshape (c′1+c′ 21 )+f reshape (c′ 22 +c′3)
[0080] p′3=f reshape(c′ 22 +c′3)
[0081] Secondly, feature fusion is performed in the spatial dimension. Through the 1*1 convolution operation, c″=[c″1, c″2, c″3] is obtained, and the upsampling operation is performed. The output p″=[p″1, p″2, p″3] is expressed as:
[0082] p1″=c″1
[0083] p″2=c″2+f upsample (c″1)
[0084] p″3=c″3+f upsample (c″2)
[0085] Finally, the feature maps of p′ and p″ are fused to obtain p = [p1, p2, p3], which is expanded along the time step direction to obtain the time series feature representation Contains information such as content historical popularity, user social relationships, and user mobility.
[0086] Based on the multi-dimensional time series features obtained by multi-scale feature fusion, this embodiment designs a multi-step popularity prediction model based on a multi-head attention mechanism:
[0087] A decoder structure based on the Bi-LSTM encoder and the multi-head attention mechanism is constructed. The Bi-LSTM layer is used in the encoder to extract historical time series features, and the output vector of the Bi-LSTM intermediate layer is input into the decoder, which contains the historical and future trends of content popularity. The popularity value of the current time step is input into the decoder, and the predicted popularity values of multiple steps in the future are output.
[0088] Bi-LSTM is used to predict the popularity of request content to extract historical and future time series features. Each Bi-LSTM neural network unit can be expressed as:
[0089]
[0090]
[0091]
[0092]
[0093]
[0094]
[0095] h t =g o *tanh(C t )
[0096] in and h t They represent the input and output vectors at time t, W represents the weight matrix of the corresponding gate, and b represents the bias vector.
[0097] like Figure 3 As shown, the vector [h t-τ ,h t-τ+1 , ..., h t ] is transmitted to the multi-head attention mechanism module, and after n fully connected layers, n subspaces are formed. Each subspace i (i∈n) independently calculates the query matrix Q i , key matrix K i and the value matrix V i , and then independently perform dot product calculations to obtain their respective attention scores:
[0098] Q i =XW i Q
[0099] K i =XW i K
[0100] V i =XW i V
[0101]
[0102] head i =Attention(Q i , K i , V i )
[0103] Then define the trainable projection matrix W o , concatenate the multi-head attention scores and project them into the output space:
[0104] MultiHead(Q,K,V)=Concat(head1,head2,...,head n )W o
[0105] Next, the output of each attention head is further processed by a feed-forward neural network, which consists of two linear transformations and a ReLU activation function.
[0106] FFN(x)=ReLU(xW1+b1)W2+b2
[0107] Among them, W1 and W2 are weight matrices, and b1 and b2 are bias parameters.
[0108] Finally, cache placement and cache replacement strategies are designed based on the predicted content popularity:
[0109] When the content reaches the node, it is weighted according to the content popularity and node distance, and the path length is divided into two segments. The node closer to the request node has a higher cache probability. The weighted popularity of the content is calculated, and the timed-out content in the local cache is cleared;
[0110] If the local cache is not full, the content, its weighted popularity, and request time are stored in the cache;
[0111] If the local cache is full, the weighted popularity of the content is compared with the least popular content in the cache:
[0112] If it is greater than the minimum popularity of the content in the cache, the minimum popularity of the content in the cache is deleted, and the new content is added to the cache to complete the cache placement and replacement strategy.
[0113] This embodiment integrates information such as the historical popularity of different contents, user social relationships, and user mobility in the social content-centric network, and designs a content popularity prediction method based on the multi-head attention mechanism. It avoids the problem of low cache utilization caused by the cache strategy that only considers the prediction of the historical popularity of the content, as well as the problem that other existing time series prediction models are highly complex and difficult to capture long-distance dependencies.
[0114] The cache placement strategy designed by the present invention is compared with four baseline strategies (History Popularity-basedPredict, DNN-based Predict, Random and Optimal), which are respectively a cache strategy that only considers the historical popularity of the content and ignores multi-dimensional features, a cache strategy based on a deep neural network, a random content cache strategy and an optimal cache strategy. The cache strategy proposed by the present invention is superior to the above four baseline strategies in terms of content popularity prediction accuracy, system average content access delay and cache hit rate, and has better cache utilization and higher user experience quality.
[0115] Embodiment 2
[0116] This embodiment discloses an active edge caching system based on multi-scale feature fusion.
[0117] Active edge caching system based on multi-scale feature fusion, including:
[0118] The integration module is configured to: construct a system network model to integrate the mobile edge computing server into the social content center network;
[0119] The multi-scale feature fusion module is configured to: based on the mobile edge computing server, respectively obtain multi-scale features representing the historical popularity of different contents, user social relationships and user mobility, and fuse the multi-scale features to obtain a unified time series feature representation;
[0120] The prediction module is configured to: input the unified time series feature representation into the encoder, extract the historical and future time series features of the content, and input the intermediate layer output vector of the encoder into the decoder, obtain the predicted popularity values of multiple time periods in the future based on the multi-head attention mechanism, and take the average value of the predicted popularity values of multiple time periods as the final prediction result;
[0121] The cache module is configured to: cache the content with the highest final prediction result as the content to be cached in advance.
[0122] Embodiment 3
[0123] The purpose of this embodiment is to provide a computer-readable storage medium.
[0124] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the active edge caching method based on multi-scale feature fusion as described in Embodiment 1 of the present disclosure.
[0125] Embodiment 4
[0126] The purpose of this embodiment is to provide an electronic device.
[0127] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the active edge caching method based on multi-scale feature fusion as described in Embodiment 1 of the present disclosure are implemented.
[0128] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.
[0129] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.
[0130] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. An active edge caching method based on multi-scale feature fusion, characterized in that: The following steps are involved: Build a system network model to integrate the mobile edge computing server into the social content center network; Based on the mobile edge computing server, multi-scale features representing the historical popularity of different contents, user social relationships and user mobility are obtained respectively, and the multi-scale features are fused to obtain a unified time series feature representation; The unified time series feature representation is input into the encoder to extract the historical and future time series features of the content. The output vector of the middle layer of the encoder is input into the decoder. The predicted popularity values for multiple time periods in the future are obtained based on the multi-head attention mechanism. The average value of the predicted popularity values for multiple time periods is taken as the final prediction result. The content with the highest ranking in the final prediction results is used as the content to be cached and cached in advance.
2. The active edge caching method based on multi-scale feature fusion according to claim 1, characterized in that: There are multiple mobile edge computing servers in the social content center network, and the multiple mobile edge computing servers are connected to each other through base station links; multiple user devices with mobility and social relationships are distributed around each mobile edge computing server; and the cloud data center provides a content library.
3. The active edge caching method based on multi-scale feature fusion according to claim 1, characterized in that: Obtain multi-scale features that characterize the historical popularity of different content, user social relationships, and user mobility, including: Define the content history popularity feature matrix to include the number of content visits in multiple time steps, user ratings of the content, and content type information, and use Bi-LSTM to extract features from the content history popularity feature matrix; Define the user social relationship feature matrix to include the user's social network activities, interaction frequency between users, and social influence information, and use graph neural network to extract features from the user social relationship feature matrix; The user mobility feature matrix is defined to include the user's location information and mobility pattern information, the user's movement trajectory data is collected, the movement distance and movement speed of each user are combined into a mobility feature vector, and the mobility feature vectors of all users are spliced into a mobility feature matrix.
4. The active edge caching method based on multi-scale feature fusion according to claim 1, characterized in that: The multi-scale features are integrated as follows: Expand the multi-scale features that characterize the historical popularity of different contents, user social relationships, and user mobility into a two-dimensional feature map; A multi-dimensional feature pyramid network is used to semantically fuse the two-dimensional feature maps in both spatial and channel dimensions to obtain a unified time series feature representation.
5. The active edge caching method based on multi-scale feature fusion according to claim 1, characterized in that: In the encoder, a BiLSTM layer is used to extract historical and future temporal features of the content; In the decoder, the encoder intermediate layer output vector is divided into multiple subspaces through multiple fully connected layers. Each subspace independently calculates the query matrix Q, key matrix K and value matrix V, and then independently obtains its own attention score through dot product calculation. The multi-head attention scores are then concatenated and projected into the output space, and finally the predicted popularity values for multiple time periods in the future are obtained.
6. The active edge caching method based on multi-scale feature fusion according to claim 1, characterized in that: The content with the highest ranking in the final prediction results is used as the content to be cached. When caching in advance, it also includes: Calculate the weighted popularity of the content based on the content popularity and node distance; Based on the weighted popularity, the content to be cached is compared with the local cache, and the cache replacement is performed on the content to be cached and the local cache.
7. The active edge caching method based on multi-scale feature fusion according to claim 6, characterized in that: When the content to be cached arrives at the node, the weighted popularity pop of the content to be cached is calculated weight ; Clear the timed-out content in the local cache; If the local cache is not full, the content to be cached will be popped weight and request time are stored in the local cache; If the local cache is full, compare the pop weight The content with the least popularity in the local cache: If the pop of the content to be cached weight If the popularity of the content is greater than the minimum popularity in the local cache, the content with the minimum popularity in the local cache is deleted and the content to be cached is added to the cache.
8. Active edge caching system based on multi-scale feature fusion, characterized in that: include: The integration module is configured to: construct a system network model to integrate the mobile edge computing server into the social content center network; The multi-scale feature fusion module is configured to: based on the mobile edge computing server, respectively obtain multi-scale features representing the historical popularity of different contents, user social relationships and user mobility, and fuse the multi-scale features to obtain a unified time series feature representation; The prediction module is configured to: input the unified time series feature representation into the encoder, extract the historical and future time series features of the content, and input the intermediate layer output vector of the encoder into the decoder, obtain the predicted popularity values of multiple time periods in the future based on the multi-head attention mechanism, and take the average value of the predicted popularity values of multiple time periods as the final prediction result; The cache module is configured to: cache the content with the highest final prediction result as the content to be cached in advance.
9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the active edge caching method based on multi-scale feature fusion as described in any one of claims 1 to 7 are implemented.
10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the active edge caching method based on multi-scale feature fusion as described in any one of claims 1 to 7 are implemented.
Citation Information
Cited By
Data cache replacement system and method based on reinforcement learning
CN119821003A
Edge computing network-oriented energy efficiency service deployment and delivery method and system
CN120751406A
