Research On Recommendation Approach Based On Deep Learning And Sea-Cloud Collaboration

With the popularity of mobile device and the improvement of Internet technolo-gies,the scale of Internet resource increase rapidly.Thus,for service providers,how to discover and recommend with massive data efficiently,and meet users' personalized needs becomes the problem worthy of study.However,in large-scale recommendation,traditional recommendation algorithms exist lots of difficulties,such as scalability,cold start,sparsity,parameters updating,and so forth,which makes it hard to provide per-sonalized service for users.This dissertation focused on the sparsity in recommendation and researched in the following three issues:1.The existence of user preferences drives the user to interact only with items of interest in the recommender system,which leads to data sparsity.At the same time,user preferences will change dynamically over time.Thus,the sequential feature of user preferences is worthy of further study.2.The abundance of network resources increases the scale of data in recommender system,makes the data more sparse,and highlights the imbalance of data distri-bution.How to learn the features of rare items,solve the problem of data sparsity,and meet the individual needs of users are the hotpots and difficulties in the re-search of large-scale recommendation.3.The growth of mobile users and the increase of users' personalized requirements have brought huge load pressure to the existing recommendation systems.How to offload the requests of the cloud-based large-scale recommendation system,meet the individual needs of users is the key point of personalized multimedia recommendation research recently.To solve these problems,in this dissertation the main research work and innova-tions are focused on four parts:1.To investigate the sequential features of user preference,we studied the user pref-erence based on item similarity,and proposed a recommendation algorithm based on the sequential features and deep neural network.We divide the user's sequen-tial data into different slices according to the timestamp.Then we calculate the similarity of items between and in the slices to analyze the sequential features of user preference.Based on the sequential features,we constructed the recom-mendation algorithm based on the deep recurrent neural network.And based on the top-N recommendation and sequence recommendation,we improved the cost function of the neural network and the generation method of the recommendation list to balance and optimize the two recommended results.Experimental results show that the hit rate of top-10 recommendation list generated by the improved recommendation algorithm can reach 78.6%.2.To reduce the sparsity of data in large-scale recommendation system,we proposed a recommendation algorithm based on the shared embedding vectors.We allo-cated an item table for items,items on the same row or column are represented by a shared embedding vector,thus,each item in the table can be presented by a pair of embedding vectors.Then,we constructed the recommendation algo-rithm based on the deep recurrent neural network,and presented a initialization method and an adjustment method for the item table based on the item similarity and graph.We also improved the loss function to speed up the training proce-dure.Experimental results show that the proposed recommendation algorithm can improve the hit performance and handle the new user cold start problem.3.To offload the user requests,and meet the individual needs for users,we proposed a dual-neural-network-based framework for distributed personalization.The pro-posed framework includes a global model and several locally personalized mod-els,which is similar to the sea-cloud collaboration framework.The global model,which is placed on the cloud,learns users common patterns from global data and provides the common service for new users.The personalized model,which is placed on the edge of the network,learns users individual features from indi-vidual data based on the global model and provides the personalized service for users.In order to decrease the communication cost between end users and the central server and protect data privacy,we presented to use two neural networks to implement the global and personalized recommendation model,which named as memory module and response module.Experimental results show that our framework improves the recommendation performance,the personalized model has 5.4%of accuracy than the global model.4.To provide personalized service for numerous users,based on the sea-cloud col-.laboration framework,we proposed a personalized multimedia service system based on the smart routers and edge devices.The system takes full advantage of the edge devices to relieve the load of content delivery network.The routers and edge devices are used for the deployment and cache of video resources.Fur-.thermore,the edge devices can provide part of local services for users instead of content delivery network.We also presented a hybrid node management algo-rithm for large-scale,flexible,high churn network to manage the routers and edge devices.The evaluation results show that up to half of the user requests can be serviced by the proposed system without accessing the content delivery network,which means the proposed framework can distribute the load of cloud to the edge device effectively,and provide real-time response for users locally.