Research On Recommendation System Based On Deep Neural Network

With the rapid development of information technology and Internet,the network application has developed from a content-centered information provider to a humancentered interactive network platform.Nowadays,we are entering the era of big data brought about by mobile Internet.Under the background of big data era,numerous Internet applications have brought great convenience to people's lives and entertainment,but also caused some problems.On the one hand,the massive information makes information overloading become a normal phenomenon,and Internet users encounter great trouble when they are looking for information related to their interests.On the other hand,how to present massive information to potential interested users on the Internet platform to improve service quality is also a great difficulty for service providers.Recommendation system is designed to solve these problems.Its main function is to find the content that users are interested in from the huge amount of information.The successful practices of recommendation system in taobao,amazon,Youtube,and other Internet applications prove that recommendation system can not only bring good experience to Internet users,but also bring huge business value to enterprises.At present,the traditional recommendation methods are mainly divided into contentbased recommendation methods and collaborative filtering based recommendation methods.Content-based recommendation only screens out those similar items for users through text attribute features,it lacks mining of users' feedback information,and has problems such as homogenization of recommendation content and cold start.Recommendation methods based on collaborative filtering rely on the interactive information between users and items,and suffer from problems such as data sparsity.Based on the above situation,this paper studies the application of deep neural networks in the recommendation problem.We explore and attempt to solve the problem that the traditional recommendation method can not well be applied in the specific Internet application scenarios.The main research contents are as follows:1.Aiming at the scenario of e-commerce,this paper takes session recommendation problem as the research target and proposes a session recommendation model based on attention mechanism and neural networks(UserVec-Att).On the one hand,we consider the long-term preference of the user and model vectorization representation of user's session by utilizing the multi-layer perceptions.On the other hand,considering the particularity of the session recommendation problem,the user's short-term dynamic interest is an important influence factor in the user's next click behavior,so the attention mechanism is adopted to model the user's session behavior sequence and construct the user's short-term interest representation.Our model models the user's next click behavior as a multi-classification problem and provides Top-K recommendation.Experimental results show that our model has better recommendation performance and the effectiveness of attention mechanism in processing session sequence problem.2.Aiming at the scenario of location-based social networks,this paper selects the point-of-interest(POI)recommendation as the research target and proposes a point-ofinterest recommendation model based on graph embedding representation and neural networks(SG-NeuRec).In this paper,the users' social graph and the POIs' geographical graph are abstracted in the form of unweighted graph.By introducing an unsupervised learning method based on autoencoder,the graph embedding representations of users and POIs are obtained.Then,we learned the interaction representation between users and POIs by using deep neural networks,and integrated social,geographical,temporal influence factors under a unified neural networks framework.Experimental results show that our model effectively improves the accuracy of POI recommendation.