Research On Recommendation Algorithms Based On Improved Variational Autoencoder

With the increasing amounts of online information,recommender system has been playing more indispensable role in helping people deal with information overload,and boosting sales for e-commerce.Recommendation algorithm obtains the potential relationship between users and items by analyzing the historical behavior data of users,predicts users' preferences for items and makes recommendations.However,the traditional recommendation algorithm has many shortcomings,such as: can not handle high sparse data well,cold start problems for new users or new items,can not better understand the characteristics of users and items,poor scalability and so on.The recommendation algorithm based on variational autoencoder is different from the traditional recommendation algorithm.Most of them have good extensibility,and can also understand the user's needs,item characteristics and the relationship between users and items.In view of the above performance,the application of variational autoencoder in the recommendation filed has attracted wide attention.Recommendation algorithms based on variational autoencoder are studied,and two improved algorithms are proposed.(1)Recommendation algorithm based on denoising variational autoencoder.The recommendation algorithm uses training data superimposed with noise when training the model.Because noise is unavoidable in actual test data,the algorithm can learn data characteristics and noise characteristics when using training data with noise.In this way,the new algorithm can be made more robust and the recommendation quality is improved.In order to further improve the recommended performance,the multinomial likelihood function and variational inference are used to tune the model parameters in the new algorithm.(2)Recommendation algorithm based on adversarial variational autoencoder.Denoising variational autoencoder uses the idea of variational inference in the process of implementation,which leads to some differences between the original data and the data generated by it.In response to this situation,adversarial variational autoencoder is proposed and applied to the recommendation.The adversarial variational autoencoder uses the adversarial ideas in generative adversarial network,and denoising variational autoencoder is used as the generation model.In this way,the reconstruct capability of model is improved,and the robustness and accuracy of the recommendation are improved.Firstly,the principles of the proposed two recommendation algorithms are elaborated in detail from the theoretical aspect,and their feasibility is analyzed.Then,corresponding comparative experiments are designed for the two algorithms.The experimental results show that the recommended performance of the two improved algorithms is better than the comparative algorithm.