Research On Application Of Autoencoder In Recommendation System

The autoencoder model is important in the field of machine learning,which is widely used in many fields and shows amazing results.This dissertation mainly focuses on several autoencoder models suitable for recommendation systems:classical autoencoder,graph autoencoder and variational autoencoder.According to the application requirements of personalized recommendation system,this paper transforms these three types of autoencoder models and combines them with other recommendation technologies to improve the performance of existing recommendation systems.The main research work and innovation of this thesis is summarized as follows:(1)According to the structural characteristics of graph autoencoder,this thesis proposes a rating prediction method based on graph autoencoder combined with social influence model.For the message transmission mode in the graph structure,the graph encoder has a good feature extraction ability.In this method,the user-item bipartite graph is constructed,and the rating prediction task is turned into a link prediction problem,which is reasonably combined with the global influence modeling.In addition,the interpretability of the model is also analyzed.In this dissertation,a large number of experiments have been done on public data sets,and the experimental results show that the prediction method proposed in this paper is obviously superior to the existing related methods.(2)Applying the compression coding paradigm of classical autoencoder to the financial field,a lightweight non-profit crowdfunding project recommendation framework is proposed.Different from the existing methods aimed at improving the return on investment,this recommendation framework aims to improving the investment turnover rate of public welfare crowdfunding projects.Experiments show that the prediction accuracy of the recommendation method based on this framework reaches the most advanced performance in this field.(3)In this thesis,the hidden space structure of variational auto-encoder is studied,and a hybrid recommendation model of balanced encoder and decoder is proposed.First of all,for the posterior collapse of the recommendation model of variational auto-encoder in the training and optimization process,the reason is the mismatch between the encoder with poor optimization and the decoder with strong representation ability,that is,the poor mapping from data manifold to parameterized graph,which makes it difficult to learn the transformation graph between them.In order to solve this problem,this paper transforms the hidden space of encoder,and proposes a recommendation model that can effectively balance encoder and decoder.Experimental results show that the scheme can effectively alleviate the posterior collapse problem.