Research On Edge Caching Technology Based On Content Popularity Prediction

With the rapid development of the mobile Internet,short video platforms represented by Bili Bili and Douyin have risen rapidly.Most of the short videos are released by users,their release is random,and they get a lot of visits in a short period of time once they are released.Repeated access to popular short videos brings huge load to the core network.Cache popular content to the edge of the network closer to users,which can reduce network load and reduce user access latency.How to quickly and accurately predict the popularity of short videos is the key to efficient caching.Most traditional content popularity prediction algorithms use early popularity to predict future popularity for text,images,and long videos,ignoring the characteristics of short videos.To this end,this thesis proposes a short video content popularity prediction algorithm based on multimodal features.In the early stage of short video release,it only predicts the popularity level of short video content based on the publisher's portrait and short video text characteristics.When a short video is released for a period of time,predict the content popularity of multiple time slots in the future according to the publisher's portrait,short video text and social characteristics,and early playback volume.On the basis of predicting the popularity of content,this thesis proposes an edge caching algorithm based on deep reinforcement learning to reduce user access delay.Therefore,this thesis focuses on content popularity prediction and edge caching strategy.The main research contents are summarized as follows:(1)In view of the inability to collect enough historical data for accurate prediction of content popularity in the early stage of short video release,this thesis proposes a Content Popularity Rank Prediction based on Convolutional Neural Networks(CPRPC).The model only predicts the playback level based on the uploader feature and text feature of the short video.First,multimodal deep features are extracted using convolutional neural networks.Then,popularity classification is performed using the Softmax classifier.Finally,the effect of different modalities on the classification of content popularity is analyzed.The simulation results show that the proposed model can effectively improve the classification performance of content popularity.(2)After the content is published for a period of time,this thesis proposes a Content Popularity Prediction based on Multimodal features with Explicit Duration Recurrent Networks(CPPME).Predict the content popularity of multiple time slots in the future based on UP portraits,short video text and social characteristics,and early playback volume.Convolutional neural networks are used to construct multimodal feature vectors of short videos,and the multimodal feature vectors are used as the initial cell state of an explicit duration recurrent neural network to capture the influence of multimodal features on content popularity.The CNN and the EDRN are jointly trained,and the initial cell state of the EDRN is updated through the forward propagation algorithm,which is used as the label of the multimodal external feature,and then the CNN is trained.Compared with unsupervised multimodal feature learning models,joint training can improve the fit between multimodal features and content popularity,thereby improving the prediction accuracy of the model.The experimental results show that,compared with other content popularity prediction models,the CPPME model proposed in this thesis has better prediction performance.(3)On the basis of content popularity prediction,this thesis proposes a Deep Reinforcement Learning for Edge Caching(DRLEC).According to the prediction results of content popularity,the newly released short videos are cached and the cached short videos are replaced to improve the cache hit rate and reduce the download delay.The experimental results show that,compared with other caching strategies,the edge buffer replacement strategy proposed in this thesis can effectively improve the cache efficiency.