Research On Transfer Learning Based On Feature Reconstruction And Sample Selection

With the rapid development of the big data industry,more and more industries need to conduct in-depth analysis of relevant data in order to mine useful information from the data,which helps to establish the model for the right domain task and make great benefits.In most of the business processes based on machine learning,data is always the core.With good quality data,a better prediction model for the task can be established.However,there are still some thorny problems in traditional machine learning.For example,in order to build a good model,it's necessary to collect amount of labeled sample data,but some data is very difficult to obtain,and it may take a lot of money,material and financial resources to obtain the data from that domain,and sometimes it is even impossible to access to the data in certain fields.As a new research field,transfer learning can transfer knowledge learned in one domain to a different but related domain.In the field of transfer learning,how to effectively adapt the data from different domains and how to filter out the effective samples in the source domain to assist the training for the target domain-modelling are-the two main challenges.The paper firstly analyzes both the feature-representation-based transfer learning algorithms and the instance-based transfer learning algorithms,designs and verifies three kinds of cross-domain transfer learning methods.Firstly,this paper proposes a simple instance-based transfer learning model.Through a non-iterative sample selection work on the source domain data,the model can enrich training samples of the target domain and train a transfer learning model quickly.Experiments show that the simple transfer model on the cross-domain facial expression data sets can have better prediction results than those without the transferring method.In addition,this paper discusses the multi-feature ensemble transfer learning method based on mutual information weighting,which can learn the information in the source domain under multiple features through multiple perspectives.Then,this paper innovatively introduces mutual information to measure those transfer works under different features.After each data similarity of both domains of each feature expression is calculated,the learning results under multi-feature expressions are ensembled by exponential weighting method.Experiments show that this method can effectively ensemble those weak transfer learners' performances under multiple different features and achieve better performance than single feature tansfer learning method.In order to further improve the transfer performance,this paper introduces the auto encoder to adapt data from different domains which can help establish the relationship between the source domain and the target domain.With the help of mutual information which guide the training in a deep neural network,this method can map the source domain data to the target domain,as well as maintain the minimum data distribution difference between the source domain and the target domain.Finally,by using the simple transfer model again,a more universal transfer learning method based on feature reconstruction and instance selection is obtained.