Content-driven Edge Network Proactive Caching Method

With the increasing data requests and content access from users,the surging data traffic poses great challenges to network capacity and backhaul links.Researchers have tried to solve the problem by enhancing the performance of network equipment and deploying more base stations,etc.Although they have achieved certain results,they still cannot support the long-term evolution of network performance.To cope with this problem,the edge network caching technology arises at the right moment.It caches the contents that users may access to the edge devices closer to users,so that users do not need to go through the backhaul link when requesting these contents,thus liberating the pressure of the backhaul link and improving the network performance.However,the storage capacity of edge devices is often limited and cannot store all the content of the Internet,so it is necessary to predict the content that users may request in the future and cache it in advance,which is the main work of caching method.Although a large amount of research on caching methods has been conducted at home and abroad,there are still problems such as low prediction accuracy and poor network performance,mainly due to the lack of more detailed consideration of data,which are specifically shown in:(1)For the scenarios with large user changes and high fluidity,user changes rapidly over a period of time,and user's behavior data is relatively sparse.Therefore,it is not meaningful to mine the preferences of users in the last time period to predict the future popularity of content.The focus is on the correlation between content,but current researches do not fully explore the features contained in the content,and most of them focus on the prediction effect of the model,ignoring the running time;(2)For the scenarios with small user changes and low fluidity,user changes slowly over a period of time,and user's behavior data is dense and rich in user preference information.The user's preferences will also affect the popularity of the content.Therefore,the focus is on the mining of user preferences,but current researches often consider users' short-term or long-term preference in isolation,and ignore the influence of the fusion of long-term and short-term preference on user preferences modeling.At the same time,considering the sequential characteristics of users' requests at different moments,the dynamic changes of user preferences can be captured more accurately,and existing work often overlooks this important feature in modeling.In response to the above problems,this thesis has launched an in-depth study on the proactive caching method of edge network.The main contents of this thesis are shown below:Aiming at the problems of insufficient content feature mining and long running time,this thesis proposes a caching method based on local and global features of content called CMLGF.This method uses a convolutional neural network to model the local features of the content,and introduces a self-attention mechanism to capture the global features of the content,that is,the dependencies between various content.Then,the features extracted from the two are merged through a fully connected network,and finally outputs the probability of each content being requested,and the cached content is determined according to the probability.The experimental results based on the real dataset verify the effectiveness of the model in improving the accuracy of prediction.At the same time,the comparative experimental results show that the method also has a good performance in reducing the running time.Aiming at the problems of incomplete consideration of user preferences and neglect of sequential characteristics,this thesis proposes a caching method based on the fusion of user long-term and short-term preference called CMLSP.This method considers both the longterm and short-term preference of users.The long-term preference of users is modeled through metric learning,and the short-term preference of users is modeled by self-attention mechanism.In addition,sequential characteristics are integrated into the short-term preference modeling process,making the model sequence-aware and stronger ability to capture the dynamic variation of the user preferences.Then,the user's long-term preference and short-term preference are combined through linear transformation.Finally,each content is scored by measuring the distance from the long-term preference representation and the distance from the short-term preference representation to get the final result.Experimental results based on real dataset verify the effectiveness of the model in reducing network load and improving user satisfaction.