Research On Semi-supervised Learning With Ladder Network Method And Application

Semi-supervised Learning is a hotspot research issue in the field of Deep Learning,which combines information from both large number of unlabeled data and limited amount of labeled data.As Semi-supervised Learning has an obvious advantage in the aspect of improving performance and reducing the cost of artificial annotation,it has been applied to solve practical issues.Laddar network is a new network structure in deep learning fields.In this dissertation,the research on the structure of laddar network is investigated and the advantage of laddar network in dealing with semi-supervised learnig is analyzed.After that a semi-supervised learning with laddar network is proposed and applied on recommender systems.The main contributions of this dissertation is elaborated as follows:(1)The laddar network is introduced and the detailed learing process of laddar nework has been deduced.Adding a lateral connection between the encoder and decoder on each layer leads to the laddar network.learning rules of the laddar network have been deduced and the advantages of usage for semi-supervised learning have been analyzed(2)A semi-supervised learning method with laddar network is proposed.Based on the improvement of the laddar networks,a framework of semi-supervised learning with laddar network is proposed,and the steps of the implements of the framework have been designed,including building the encoder,decoder and the cost funtion.(3)Proposed the modeling method of recommender systems based on semi-supervised learning.In the term of the common recommender systems,including the Content Based Recommender System,UserCF,ItemCF and the Label Based Recommender System,the data models are reconstructed.So that recommender systems are compatible with semi-supervised learning methods,which enables the recommender systems to deal with unlabeled data.For the validation of the proposed method,experiments on MNIST,CIFAR-10 and missile intercept data have been conducted.And the Semi-supervised Recommender Systems have been used for movies recommendation.The results turn out that the proposed method is feasible.