Research On Collaborative Filtering Recommendation Algorithm Based On Generative Adversarial Networks

With the rapid development of technology,the information on the Internet is increasing exponentially.It is difficult for users to quickly and accurately associate the information they need from these massive amounts of information,which leads to information overload.The recommendation system can help users filter information and is an effective way to alleviate information overload.The recommendation algorithm is the core component of the recommendation system,and it directly affects the performance of the recommendation system.Traditional recommendation algorithms have some shortcomings.They cannot learn the highlevel interaction characteristics between users and items well,and they also face problems such as cold start and data sparseness.Due to data sparseness and insufficient learning ability of the model itself,some existing recommendation algorithms still have the problem of insufficient user item interaction information mining and inability to learn the user's preferences well,which results in the recommendation accuracy not reaching a satisfactory level.In recent years,the Generative Adversarial Networks have made breakthrough achievements in many fields.By learning the distribution of input data,it can generate a distribution consistent with the input data.Through the study of the distribution of user item interaction data,the user preference information for all items can be obtained,which provides a new research idea for the research of recommendation system.This dissertation studies the recommendation algorithms at home and abroad,focus on the recommendation algorithms based on Generative Adversarial Networks,which mainly includes the mining of user item interaction information,the learning of user preference information,accuracy improvement and so on.The main work of this dissertation includes:1.In view of the insufficient number of user item interactions accompanied by data sparsity in the existing recommendation system,and the inability to learn user preference information well to reduce the recommendation accuracy,this dissertation propose an adversarial collaborative filtering recommendation algorithm combines positive and negative items reconstruction losses(PNACF).The algorithm learns the actual interactions of existing user and items by Generative Adversarial Networks and adds missing interaction information,thereby generating user preference for all items,to a certain extent,alleviating the problems caused by data sparseness;At the same time,this dissertation add the penalty of the positive and negative items reconstruction losses to the objective function of the generative model,which promotes the model's ability to mine user item interaction information,and effectively improves the recommendation accuracy.This dissertation uses two datasets for experiments,and the experimental results verify the effectiveness of the PNACF algorithm.2.In view of the insufficient ability of existing Denoising Autoencoder to learn user item interaction information,and the inability to learn user preference information well,which leads to the problem of low recommendation accuracy,this dissertation propose a collaborative filtering recommendation algorithm based on Adversarial Denoising Autoencoder(ADAE).The main idea is to introduce a discriminative model on the basis of the existing Denoising Autoencoder,so that the reconstructed user preference vector not only meets the requirements of self-reconstruction,but also meets the requirements of the discriminative model,so as to obtain the more accurate user preference information to improve the recommendation effect.Proved by a series of related comparative experiments,the learning ability of the model can indeed be effectively improved by introducing the discriminative model,better user preference information can be learned,and the recommendation accuracy can be finally improved.