A Research On Recommendation Algorithm With Deep Feature Based On Deep Learning

The rapid development of mobile communication network,it is the key to enhance user experience to select the items that attracts users from the massive items.Recommendation algorithms have been widely studied and can provide users with personalized decision support and information services.In recent years,the deep learning based recommendation has been researched a lot,which usually has better performance than the traditional recommendation algorithm.The hybrid recommendation algorithm integrates the auxiliary information into the traditional recommendation algorithm to solve the data sparsity problem commonly faced in recommendation.However,the current utilization is not effective enough,and the underlying characteristics are ignored.In addition,the review contains a wealth of user preferences and item characteristics,and the existing review-based recommendation algorithms face many problems,such as information loss,noise,and loss of the corresponding relationship between ratings and reviews.Hence,this thesis carried out indepth research,extracted the deep features under the framework of deep learning,proposed a convolutional collaborative filtering network,a new recommendation algorithm exploiting reviews and a recommendation algorithm simultaneously modeling reviews and ratings.The main work of this thesis are as follows:1.In this thesis,the hybrid recommendation algorithm with additional information and the recommendation algorithm with reviews are researched in depth,and the shortcomings of the existing research are analyzed.2.A convolutional collaborative filtering network(CCFNets)is proposed.The current advanced hybrid recommendation uses auxiliary information in a limited way and the linear modeling method limits the recommendation performance.Based on the deep learning framework,CCFNets analyzed a variety of relationship,and finally constructed the non-linear shallow features,further explored the deep features of interaction,and extended the use of auxiliary information.Experiments on real datasets show that CCFNets are superior to the most advanced hybrid recommendation algorithms,especially in the case of sparse data.3.A new recommendation algorithm exploiting reviews(RRNets)is proposed.Most of the advanced review-based recommendation algorithms adopts review splicing,which causes problems such as information loss and ignores the deep interactive features.Considering the particularity that reviews after interaction,this thesis proposes a new way to use reviews,constructs the user's review representation of non-interactive items,and effectively promotes the rating prediction.The algorithm can also optimize the presentation of users and items through implicit feedback,thus improving the ability of review construction.Experiments on real datasets show that RRNets is superior to the most advanced one which exploit reviews,and the promotion of implicit feedback on the algorithm is verified.4.Two methods are proposed to improve RRNets,based on which a new recommendation algorithm IRRNets is proposed to simultaneously model reviews and ratings.The algorithm further considers the difference between the rating domain and the review domain and uses the idea of transfer learning for reference to solve the problem of domain adaptation.IRRNets are highly scalable and can form a mutual promotion network by combining the collaborative filtering,while optimizing the accuracy of rating prediction and review construction.Experiments on real datasets verify that IRRNets can further improve the performance of RRNets and the promotion of each improvement point.