Research On Collaborative Filtering Based On Wasserstein Autoencoder

Instead of the era of information,this world is already moved into the era of data.Using a large amount of information in the data age can bring a lot of convenience to our life,but how to use this information has become a big problem.The recommendation system can help users quickly locate the target information or items through technical means and make full use of the information.However,the essence of traditional collaborative filtering algorithms,such as potential factor model,is linear model,which fails to solve the problem of sparse data and the limitation of linear model.Therefore,it is needed to propose deep neural network models to resolve recommendation issues.This thesis is thus proposed with the purpose of building an effective nonlinear model for collaborative filtering to make recommendations from large-scale and sparse data.Aiming at this research issue,a novel Wasserstein based auto-encoder model is proposed in this thesis.Based on the structural characteristics of the new auto-encoder,the content of the proposed approach includes model design,a neural network component which can capture hidden nonlinear features.By introducing the sparse regularization term,we also redesign the distribution distance of implicit variables and redesign the reconstruction error loss function.Finally,the proposed model is applied to the implicit feedback experiment.The proposed model is believed to be able to alleviate the sparsity issue of recommendation system.Moreover,this thesis further considers the correlation among users.Thus,this thesis extends the previous model to take into consideration the correlation of user ratings.In this model,we begin with simple undirected acyclic graph and derive a series of formula.On top of this,we generalize the case of acyclic graph,and optimize the newly proposed model.A new auto-encoder model is designed by integrating data correlation.At last,we evaluate the proposed two models as well as other state-of-the-art algorithms on several large-scale and sparse data sets.The experimental results demonstrate that the proposed models are superior to the compared methods with respect to prediction accuracy.This verifies the effectiveness of proposed approach.