Research On Recommendation Algorithm Based On Deep Learning

In recent years,due to the rapid development of big data and cloud computing technology,the amount of data such as sound,image and text shows an exponential growth,and the problem of information overload is getting worse and worse.How to help users to recommend the information they need from the massive data has become an urgent problem to be improved,so the recommendation system is getting more and more people's attention.Sequence recommendation is a very important task in the recommendation system.It regards the interaction between users and items as a dynamic sequence,models the sequence of items by users,and predicts the next behavior according to the changes of users' interests.Due to the continuous development of computer hardware and deep learning technology,recommendation algorithm based on deep learning has become a research hotspot.In this paper,the convolutional neural network method is applied to the recommendation algorithm,focusing on the sequential recommendation problem.By exploring user behavior patterns and capturing the characteristics of sequence data,the main research work is as follows:1.Caser,based on the deep learning algorithm,maps the user's project interaction sequence to the embedded space and uses the convolution filter to learn the basic features of the sequence pattern,which can effectively capture the user behavior of single point,joint and jump types.However,for some discrete user behaviors,it is not good to handle the case that items have intervals.Because multiple projects in the recommended item has a great advantage,which can effectively deal with discrete problems of the project, can from different angles such as metadata,the function of fine-grained understanding of the project,so this article will be more projects Caser combination of relations and based on the deep learning,puts forward the embedded sequence of personalized recommendation algorithm based on convolution sequence(CICR),When items are discontinuous,better recommendation effect can be achieved by integrating item relationship model.2.The CICR algorithm uses a vertical filter,which aims to generate a weighted sum of all previous items,but it only performs summations along each dimension and lacks representative channel interactions.This weighted sum applies only to shallow one-layer network structures,leading to limitations when modeling remote dependencies or large-scale data flows that require deeper architecture.Therefore,in this chapter,pairwise coding is used to deepen the number of network layers,and multi-item relations are fused to encode each behavior sequence into high-dimensional features,which are then applied to the convolution layer.In this method,the image features of the project are input into 2D convolutional neural network,which makes the channel interaction between vectors,and further makes the network flexibly adapt to the shallow or deep structure of different tasks by filling operation.Experiments are carried out on Hetrec2011,a public data set,and Tourism,a data set of tourist attractions constructed based on "2018 Cloud Transfer Cup National Tourism Big Data Challenge" and "Whale Community Tourist Attractions Data".The experimental results effectively prove that the proposed algorithm has better recommendation performance than other popular recommendation algorithms,and verify the rationality of the proposed algorithm.