Research On Personalized Recommendation System Based On Deep Learning

In the last few years,with the accelerated popularization and rapid evolvement of Internet technology,various types of information have shown explosive growth,which makes it difficult for users to find the information they need in the mountain of information.In order to help user find the content they need quickly and effectively,therefore,the recommendation system came into being.In the process of surfing the Internet,users leave a large amount of implicit behavior information,which is of great significance for exploring users' preferences.At present,many researchers have made some researches on recommendation algorithms integrating implicit feedback,which can be divided into two categories: one is to establish user interest and rating based on implicit feedback data;The other is the improvement of matrix decomposition technique.Therefore,this thesis also makes improvements for these two categories.For the first category,users' preferences are converted into multiple classification problems based on implicit feedback data,and using deep learning technology to fully explore users' implicit characteristics;for the second category,implicit feedback information can solve the problem of poor recommendation effect caused by the sparsity of the rating.By optimizing the similarity of existing users and items,this thesis tries to make full use of social and tag information to improve recommendation performance.The specific work and innovation of this thesis is as follows:(1)In view of the diversity of web browsing behavior and the characteristics of different interests and preferences among users,a user preference classification model based on web browsing behavior is proposed.First of all,the convolutional neural network is used here to extract the features of various browsing behaviors locally.Secondly,capsule network was used to extract the overall behavior characteristics.Finally,the softmax classifier is used to predict users' browsing preferences.Experiments show that the model can accurately predict users' emotional preferences.(2)In order to solve the problem of poor recommendation effect caused by the inaccuracy of similarity between users and items based on rating in traditional recommendation system,a probabilistic matrix factorization recommendation model based on social trust and tag semantic similarity is proposed.Firstly,the users' similarity is improved by constructing a new social trust to describe the relationship between users.The new social trust is built by using users' friend relationships,preference similarity and tagging behaviors.Then,the semantic similarity between item tags is calculated by Sentence-BERT,the tag pair with the greatest similarity is selected successively,and the weight of each tag is added,so as to calculate the similarity between items.Finally,in the probability matrix factorization process,social network trust and project similarity are integrated into the implicit rating data,and the potential feature vectors of users and items are calculated to predict the rating and make Top-N recommendations.The proposed model is validated on last.fm-2k dataset,and the results show that the proposed model can effectively improve the sparse and cold start problems of the recommendation system.