Research On Ensemble Learning Based On Truth Discovery

Ensemble learning is a very important and practical machine learning method,but Ensemble learning is not a specific machine learning method,it is to generate and combine multiple base learning algorithms to complete the task.Ensemble learning effectively promotes the development of information fusion,data modeling and data mining etc.The combination of base learners is an important research topic in the ensemble learning.The existing conbination methods of base learners can be roughly divided into three kinds:average method,voting method and learning method.the voting method is the most widely used in ensemble classification.There are many kinds of voting methods,which can be roughly divided into weighted voting and non weighted voting.In the past,the voting methods for the combination of basic learners in ensemble learning are too simple and inefficient.Therefore,this dissertation studies the integration strategy of ensemble learning.The main research work is as follows:（1）A new heterogeneous ensemble classification method ECTD-S based on truth discovery is proposed.In the process of calculation,the method first initializes the weight of the base classifier,i.e.the credibility of the base classifier,and then infers the credibility of the classification according to the inference rules found by the truth value,and then iteratively updates the credibility of the base classifier according to the inference rules.When the convergence condition is reached,the algorithm stops iteration and gets the final prediction result.The experimental results show that ECTD-S is superior to the comparison algorithm in Precision,Recall and F₁,which proves the effectiveness of the method in heterogeneous ensemble classification.（2）A new ensemble classification method ECTD-B based on truth discovery is proposed.The traditional Bootstrap based bagging homogeneous ensemble classification method can not dynamically update the weight of the base classifier.ECTD-B first does bootstrap sampling for the original training set to get multiple training subsets,then gets multiple homogeneous base classifiers based on these training,and then forecasts the test set to get preliminary prediction results.The next calculation process is the same as that of ECTD-S.Firstly,the weight of the base classifier is initialized,then the credibility of the base classifier is obtained based on the inference rules of truth value discovery,and then the credibility of the base classifier is updated according to the inference rules iteratively.When the algorithm converges,the final prediction results are obtained.This method can dynamically set the weight of the base classifier in the prediction stage,experimental results show that this method can also achieve better results in the field of homogeneous integrated classification.