Research On Hybrid Recommendation Algorithm Based On Variational Autoencoder

With the continuous development of Internet information technology,the volume,speed and complexity of online information are increasing rapidly.Personalized recommendation system has become an important way of extracting effective information from tedious complicated online data,which is widely used in various fields of industry.Traditional methods such as Matrix Factorization based Collaborative Filtering have proven to be very effective.Although implicit feedbacks such as browsing and clicking are much easier to collect than explicit feedbacks like movie ratings,product reviews etc.Such methods do not perform well due to the sparsity of features and the impact of so-called cold-start problem.Meanwhile,deep learning methods have drawn much researchers' attention for their huge success in the field of image classification and natural language processing which proves its excellent performance in feature processing.Hence,applying deep learning methods to recommendation system has become an important direction in this field.However,the existing models based on deep neural networks tend to deal with the content characteristics of users and items instead of capturing the interaction between users and items which is the key problem of recommendation systems,and most related works focus on using latent embeddings generated by matrix factorization frameworks.In this thesis,we propose a hybrid recommendation algorithm based on variational autoencoder in which neural network is used to capture the interaction between users and items,and generalized matrix factorization is used to guide the optimization goal of neural network and provides explanation.The main contributions of this thesis are listed as follows:First,this thesis applies variational autoencoder in recommendation systems to model interaction and content features of users and items.By encoding sparse feature using Factorization Machine,we perform an automatic high order feature combination to alleviate the problem of sparse features.Moreover,we fuse the multi-view data features of users and items into the variational autoencoder to solve cold start problem.Second,this thesis provides an explanation for the generation process of latent vector through the variational inference in autoencoder.By assuming the probability distribution limits,we proved that it's necessary to extend the inner product operation of latent vectors to a generalized matrix factorization.Third,we conducted comprehensive experiments on well-known public real datasets and compared our approach with the most widely used and most classical and state-of-the-art works to illustrate the effectiveness of our model.