Data Stream Classification Research Based On Transfer Learning

Data streams classification,an important branch in the domain of data mining,can obtain valuable information from data streams,and it has become one of the research hotspots.Data streams owns variable,high volume and high velocitye feature,and these traits lead to traditional classification methods are faced with many difficulties,such as sample labeling and concept drift.Therefore,how to establish accurate classification models and detect the change of data streams in real time have become the challenge of data streams classification.Transfer learning can transfer knowledge from the source domain to the target domain by making use of the similarity between the source domain and the target domain.Data streams classification method based on transfer learning can construct a efficient classification model that do with the problems of sample labeling and concept drift in data streams classification.Therefore,data streams classification research based on transfer learning has important practical significance and research value.In view of this,this dissertation applies transfer learning to data streams classification,and the specific work is as follows:1.Firstly,the basic knowledge of data streams classification technology based on transfer learning are explored in detail,then the common methods of data streams classification based on transfer learning are summarized.Finally,according to transfer learning,how to select the most suitable source domain classifier are introduced.2.When one depends thoroughly on the nearest neighbor's information for predicting class label of sample the wrong decision may occur,and this phenomenon is called as the pseudo nearest neighbor.In order to effectively avoid the pseudo nearest neighbor phenomenon in data streams classification,a multi-source transfer learning method based on the mutual nearest neighbor is proposed.The main idea of this method is to use the mutual nearest neighbor idea to select the mutual nearest neighbor sample set and calculate the local classification accuracy of each source domain classifier about the mutual nearest neighbor sample set.The source domain classifier with the highest local classification accuracy is combined with the target domain classifier by weighted mechanism.Simulation experimental results show that the classification accuracy and the anti-noise stability of the proposed method are significantly improved.3.In order to avoid the adverse effect of the uncertain classifier in the environment of data streams with noise,a multi-source transfer learning method based on sample certainty is proposed.This method uses the abstaining classifiers to force uncertain classifiers not to make predictions,and the core idea of this method is to calculate the sample certainty value of each source domain classifier about the target domain sample.The source domain classifier whose the sample certainty value exceeds the current threshold limit is integrated with the target domain classifier.Simulation experimental results prove that the proposed method is feasible and the classification model has high classification accuracy and good stability.This thesis explores the different selection methods of source domain classifier,and use these methods to select the most suitable source domain classifier from the source domain classifier set.The proposed methods can transfer knowledge from source domain to target domain and solve the problem of concept drift and noise in data streams.At the same time,this dissertation presents the challenges of data streams classification based on transfer learning,and discusses the future research trend of data streams classification based on transfer learning.