Research On Ensemble Of Instance-based Transfer Learning Via Multi-source Domains

With the rapid development of computer science and technology and the big data industry in recent years,data science has gradually become a new driving force for information technology and the Internet industry.Facing the explosive growth of data,how to mine the value of information in data and how to analyze the patterns implied by data information.It has become a key problem for data science to be solved.Machine learning,as a data mining method,can effectively model the data through statistical related theories and techniques.Regardless of whether it is theoretical research or practical application,machine learning has made great progress and development.However,current machine learning also has its disadvantages.Machine learning methods often require enough sample data with class labels as training data sets to establish a machine learning model in combination with statistical methods.This type of machine learning method is collectively referred to as a supervised machine.But in real reality,data in some areas may be difficult to obtain or even impossible to obtain.At this time,traditional machine learning methods will not work.Therefore,how to establish a good domain model in the absence of domain data becomes a matter of time.In the field of machine learning research,a problem of great concern was addressed.For this problem,researchers proposed the concept of practical transfer learning.As a cross-domain machine learning method,transfer learning can transfer knowledge in one domain to another one.They are related but from the different domains to model the target domain.However,traditional transfer learning often only learns prior knowledge from one source domain and migrates it to the target domain.Considering the fact that solving a specific problem in reality can actually acquire prior knowledge from multiple domains.The key problem to be solved in this learning process is to find data and find relevant sample data in the source domain and target domain.From the single source domain,the data with strong correlation in the target domain is likely to have a different data distribution.At this time,a negative transfer phenomenon will occur.That is,transfer learning will not only help the target domain to learn The better model,on the contrary,has a negative impact on the learning of the target domain.The innovation of this paper lies in the establishment of a transfer model for multi-source domains in this paper.It intends to learn implicit information in multiple domains and assist in the establishment of models in the target domain.This multi-source domain-oriented transfer learning model innovatively designs two sub-modules,namely semi-supervised learning module and integrative learning module.Among them,the introduction of semi-supervised thinking is to make better use of classless tag samples in the target domain.,and ensemble learning is to better mine useful domain knowledge from different perspectives and weight-blend its decision performance.By applying this model to Letter-recognition,20 newsgroup,and Reuters datasets and experimental testing,better prediction results than single source domain,non-integrated machine learning methods,and non-supervised machine learning methods were obtained,verifying the transfer.Learning can use prior knowledge to solve target domain tasks better than traditional machine learning.