Font Size: a A A

Research On Metallogenic Prediction Models Based On Adaboost

Posted on:2017-03-26Degree:MasterType:Thesis
Country:ChinaCandidate:S WangFull Text:PDF
GTID:2180330485991971Subject:Mineral prospecting and exploration
Abstract/Summary:PDF Full Text Request
Adaboost is an accuracy machine learning algorithm for classifying multiple kinds of patterns. The Adaboost framework can flexibly classify the data comprised by different patterns without feature selecting and overfitting problems.The main contributions of this thesis are as follows:(1) On the basis of the previous works, the metallogenic characteristics of the study area are studied by extracting different features as the dimensions of the final data according to the characteristic of the Adaboost algorithm so that multiple kinds of information can be used simultaneously.(2) An Improved Weights of Evidence Boosting method(IBoostWofE) is proposed.This method focuses on the problem of different positive and negative training error in the Adaboost algorithm by calculating the positive and negative weights of the weak classifiers respectively, instead of the previous common version.(3) An unbalance dataset robust approach has been proposed. Dataset unbalance problem is common problem in the data in mining area怂The conventional Adaboost algorithm failed to handle this problem and learned a positive-biased classifier. To solve this problem, the proposed approach weights the positive samples in the initial and finally improves the classify accuracy.(4) A cross validation approach has been proposed to solve the lack of enough training data which may reduce the generalization ability of the proposed method.(5) The conventional Adaboost algorithm and the proposed algorithm have been studied and compared.The experiments show that the initial improved Adaboost and the proposed Adaboost algorithm improve the conventional Adaboost method.
Keywords/Search Tags:Adaboost, Weights of Evidence, Improved Weights of Evidence, unbalance dataset, positive and negative error rate difference
PDF Full Text Request
Related items